RECONSIDERING VISUAL STATISTICAL LEARNING: EFFECTS OF ATTENTION, ASSESSMENT, AND ASSOCIABILTY by Kevin D. Himberger A dissertation submitted to Johns Hopkins University in conformity with the requirements for the degree of Doctor of Philosophy Baltimore, Maryland March 2021 © 2021 Kevin Himberger All Rights Reserved ii Abstract Improving the ease of learning has been a centuries-long pursuit. One tantalizing possibility is that there may be conditions under which learning occurs “automatically” and implicitly. Such a learning process would have profound impacts on pedagogy and our understanding of human development. This type of learning process has, ostensibly, been studied as “statistical learning” over the past 25 years. In this thesis, I propose that although genuine forms of automatic and implicit learning do exist, the standard paradigm for studying visual statistical learning, in adults, does not isolate an automatic and implicit learning process. Additionally, I show that some of the evidence for implicit visual statistical learning, which appeared to be strong, is in fact the result of experimental confounds in response time measures. Finally, I show that stimulus choice significantly affects visual statistical learning, suggesting links with traditionally studied associative learning. I motivate this proposal with an overview of the field of statistical learning and what distinguishes it from other forms of learning (Chapter 1). I support this proposal with evidence from four sets of experiments which investigate the: effects of automaticity and attention (Chapter 2), indirect assessment of visual statistical learning (Chapter 3), and the associability of arbitrary stimuli (Chapter 4). I conclude with a summary of our findings and how they integrate into the broader study of learning processes, along with open questions for future investigation (Chapter 5). To preview our four major findings, we find that: (1) when participants attend to stimulus motion, increasing the amount of exposure does not improve learning of visual sequence iii regularities, (2) the attentional state of participants significantly affects how well they learn statistical regularities, (3) important evidence in the literature for an implicit statistical learning process – response time facilitation within learned sequences – actually appears to arise from a methodological confound, and (4) stimulus associability is affected by stimulus choice, with faces associated more rapidly than arbitrary fractals. Collectively, these data suggest a reconsideration of visual statistical learning from an automatic, implicit learning process to a more attention-dependent and volitional learning process. Primary Reader & Advisor: Dr. Christopher Honey Secondary Reader: Dr. Janice Chen iv Acknowledgments I am immensely thankful for my advisor, Chris Honey. His tireless efforts to improve my writing, research, and presentation skills are beyond what should be expected from an advisee. Beyond science, he is truly a caring person who wants the best for people around him. Janice Chen is another excellent scientist and lovely human, who has given me more helpful feedback than some graduate students get with their own advisor. Specifically, her advice on conceptual framing and theoretical motivations helped guide my experimental design. Amy Finn provided support for the bulk of this dissertation by providing critical insight into statistical learning, both in how the subject is conceptualized it and how it can be investigated. I am lucky to have been supported by my committee members and departmental lab directors through critical feedback and suggestions on my work: Susan Courtney-Faruqee, Justin Halberda, Howard Egeth, Jason Fischer, Chaz Firestone, Brenda Rapp, and Shreesh Mysore. This work was improved by conversations with members of the Johns Hopkins Psychological & Brain Sciences department, members of the University of Toronto Psychology department, and especially with Kathrin Müsch, Feitong Yang, Eduardo Sandoval, and Buddhika Bellana. I am incredibly grateful for my parents, David & Alberta Himberger, who have supported my pursuits as if it was reasonable to leave IBM, wander the world, and then start graduate school. I am truly fortunate to have such support. I am also extremely grateful for my loving partner, Jamie. Not only has she consistently supported my goals, but also hid upstairs when the pandemic caused me to take over the dining room table. v Table of Contents Abstract ............................................................................................................................................................... ii Acknowledgments ............................................................................................................................................ iv List of Tables .................................................................................................................................................... viii List of Figures .................................................................................................................................................... ix Chapter 1: Learning from repeated exposure ............................................................................................... 1 A Brief History of Statistical Learning ................................................................................................................... 3 Standard Paradigm For Investigating Statistical Learning ............................................................................ 6 What are the defining characteristics of statistical learning? .......................................................................... 10 Automaticity ...................................................................................................................................................... 11 Implicitness ........................................................................................................................................................ 13 Artificial Grammar Learning ........................................................................................................................ 17 Serial Reaction Time Tasks .......................................................................................................................... 19 Stimulus Non-specificity .................................................................................................................................. 20 Modality-specific or domain general? ....................................................................................................... 20 What is not statistical learning? ...................................................................................................................... 22 Neural underpinnings of statistical learning ................................................................................................. 25 Prior Knowledge ........................................................................................................................................... 28 Preview of our findings ........................................................................................................................................ 29 Chapter 2: Visual statistical learning did not appear to be automatic and was strongly modulated by task................................................................................................................................................................ 31 Experiment 1: Repetition does not improve learning when task is unrelated to temporal processing ..... 33 Method .............................................................................................................................................................. 34 Procedure ...................................................................................................................................................... 36 Analysis .......................................................................................................................................................... 44 Results ................................................................................................................................................................ 47 Discussion .......................................................................................................................................................... 53 Experiment 2: Cover task modulated visual statistical learning ...................................................................... 54 Method .............................................................................................................................................................. 57 Procedure ...................................................................................................................................................... 58 Analysis .......................................................................................................................................................... 62 Results ................................................................................................................................................................ 62 Discussion .......................................................................................................................................................... 68 Chapter Discussion ............................................................................................................................................... 70 Chapter 3: Limited evidence of implicit visual statistical learning........................................................... 76 vi Experiment 3: Response time facilitation can be observed without learning ............................................... 79 The stream position effect ............................................................................................................................... 80 Motivation & Preview ...................................................................................................................................... 83 Method .............................................................................................................................................................. 84 Procedure ...................................................................................................................................................... 85 Results ................................................................................................................................................................ 86 Discussion .......................................................................................................................................................... 91 Explicit knowledge was evident in participants who exhibited the best learning ........................................ 94 Method .............................................................................................................................................................. 95 Results ................................................................................................................................................................ 96 Discussion ........................................................................................................................................................ 101 Chapter Discussion ............................................................................................................................................. 104 Chapter 4: Visual statistical learning differed by stimulus type ............................................................. 109 Why might visual statistical learning be affected by stimulus choice? ........................................................ 111 Perceptual fluency .......................................................................................................................................... 112 Depth of processing ....................................................................................................................................... 112 Prior knowledge .............................................................................................................................................. 114 Faces as a manipulation of stimulus familiarity .......................................................................................... 115 Experiment 4: Visual statistical learning was weaker when seeing abstract fractals relative to faces, but famous faces were not learned better than non-famous faces .................................................................... 117 Method ............................................................................................................................................................ 118 Procedure .................................................................................................................................................... 119 Results .............................................................................................................................................................. 120 Discussion ........................................................................................................................................................ 124 Chapter Discussion ............................................................................................................................................. 124 Chapter 5: Conclusions & Open Questions ................................................................................................ 128 Visual Statistical Learning Characteristics ........................................................................................................ 128 Automaticity .................................................................................................................................................... 129 Implicit Knowledge ......................................................................................................................................... 130 Considerations for the Target Detection task......................................................................................... 130 Explicit Knowledge ..................................................................................................................................... 131 Stimulus Non-specificity ................................................................................................................................ 133 Reconsidering Visual Statistical Learning ........................................................................................................ 134 Automaticity .................................................................................................................................................... 134 Implicitness ...................................................................................................................................................... 136 Infant statistical learning does not necessarily imply implicit & non-hippocampal learning .......... 140 Stimulus Non-specificity ................................................................................................................................ 141 Integrating with other fields .......................................................................................................................... 143 Summary .............................................................................................................................................................. 144 vii Bibliography ................................................................................................................................................... 146 Supplementary Materials.............................................................................................................................. 174 Supplemental Methods ...................................................................................................................................... 174 Supplemental Results ......................................................................................................................................... 179 Supplemental Discussion ................................................................................................................................... 185 viii List of Tables Table 1: Paradigms for studying implicit (and statistical) learning. Modified from Conway (2020). ......... 17 Table 2: Results of varying exposure (Experiment 1) ................................................................................. 51 Table 3: Descriptive statistics for Experiment 2. All p-values are Bonferroni-corrected (except rs values). The “2AFC % correct by triplets created” means are calculated only with sufficient data (4+ participants). .................................................................................................................................................................... 66 Table 4: Aspects of implicit & explicit: instruction, learning, and knowledge (content from Ellis, 2009). . 77 Table 5: Results from Experiment 3 (TD task comparison) ......................................................................... 87 ix List of Figures Figure 1: Explicit and Implicit Long-Term Memory (from Squire & Zola, PNAS 1996; Copyright 1996 National Academy of Sciences) ..................................................................................................................... 4 Figure 2: Three possible orthogonal dimensions of statistical learning (Reprinted from Conway, Neuroscience & Biobehavioral Reviews 2020). .......................................................................................... 10 Figure 3: Artificial grammar learning paradigm (Reprinted from A. S. Reber, J. Verbal Learning & Verbal Behavior 1967) ............................................................................................................................................ 18 Figure 4: Stimuli & experimental paradigm. (a) The 12 “fractal” images used. (b) Each column shows one phase of the experiment. The "Task" row indicates the participant’s goal in each phase. The "DV" row indicates the primary dependent variable measured in that phase. The experiment was designed so that more direct (more explicit) measures of VSL occurred later. Participants in the “0 Repetitions” condition skipped the “Initial Exposure” phase. ......................................................................................................... 37 Figure 5: Results from manipulating amount of repetition during the Jiggle Task. (a) Mean RTs by triplet position and condition. (b) Response time facilitation effects between triplet first and third positions. (c) Percentage of “real” triplets successfully chosen during 2AFC task. Gray dots indicate a single participant. Red dots indicate mean of distribution. (d) Mean number of triplets created (out of 4 possible). (e) Histogram showing the number of participants creating triplets, from none (0) to all (4). (f) Spearman correlation between direct measures (2AFC & Creation). Colored dots represent individual participants. Black dots indicate means (when more than four data points are available). (a-d, f) Vertical line segments indicate 95% CIs. (b-d) Dashed gray lines indicate chance performance. ........................... 52 Figure 6: Experient 2 results. (a) Performance of each subject on the 2AFC measure (% of correct choices made), by condition. (b) Histograms of the mean number of triplets successfully re-created, by condition. (c) Relationship between performance on the 2AFC task and the Triplet Creation task. rs indicates Spearman’s rank-order correlation. (d) Mean RTs by triplet position and condition. (e) Response time facilitation effects between triplet first and third positions.. Vertical line segments indicate 95% confidence intervals (where there are at least 4 data points). Dashed gray lines indicates chance performance. .................................................................................................................................. 67 Figure 7: RTs across Experiments 1 & 2 by stream position in the Target Detection Task. (a) Response times by TD stream position. Sloped lines indicate the line of best fit. (b) RTs by stream position further divided by triplet position of the target. Vertical line segments indicate 95% confidence intervals. Filled circles indicate means. ................................................................................................................................ 82 Figure 8: Target Detection Task Paradigms. (a) TD task used in Experiments 1 & 2. C2 & C3 indicate trial-specific targets. Note that the stream length dictates the triplet position of the starting element. (b) TD paradigm used in Experiment 3. "Traditional TD" indicates the standard presentation of four triplets with C2 as the example, trial-specific target. “Scrambled TD” uses the same stream position of the trial-specific target, but removes all triplet structure. Subscripts (1, 2, 3) indicate triplet position, while letters (A, B, C, D) indicate triplet identity. ............................................................................................................ 86 Figure 9: Results from Experiment 3. (a-d) Results from Target Detection (TD) Task. (e-h) Results from direct measures. (a) Response times by TD stream position of the target. (b) Same data as (a), but colored to show the interaction with triplet position. (c) Response times by triplet position. (d) Response time facilitation between first and third triplet positions. (e) 2AFC accuracy. (f) Creation task accuracy. (g) Number of participants creating each possible number of triplets (from 0-4). (h) Relationship between performance on the direct measures (2AFC & Creation). Dots indicate mean performance. x Vertical line segments indicate 95% confidence intervals. Dashed gray lines indicate chance performance. .............................................................................................................................................. 89 Figure 10: Performance for participants who reported either seeing or not seeing a pattern before starting the 2AFC task. (a) 2AFC performance by pattern seeing. Number of participants in each bin are shown (n=), as well as percentage of number of participants in that condition (%). The same values are true in a-c. (b) Creation performance by pattern seeing. (c) Response time facilitation between triplet first and third positions. Vertical line segments indicate 95% confidence intervals. Dashed gray lines indicate chance performance. .................................................................................................................... 99 Figure 11: Performance across measures of learning when separating participants who assigned labels to the images and those who did not. (a) Performance on the 2AFC task. Percentages shown in parentheses represent the percentage of participants who either did or did not use labels (same values in a-c). (b) Performance on the Creation Task. (c) Performance as assessed with response time facilitation between the first and third triplet positions. Vertical line segments indicate 95% confidence intervals. Dashed gray lines indicate chance performance. .................................................................................................. 100 Figure 12: Stimuli & performance from Exp. 4. (a-b) Matched (a) non-famous and (b) famous face stimuli. (c) 2AFC performance. Each dot indicates a single participant. (d) Creation performance. (e) Creation performance showing specific numbers of triplets created in each condition. (f) Relationship between 2AFC and Creation performance. Vertical line segments indicate 95% confidence intervals and black dots indicate group means (with n>4). Dashed gray lines indicate chance performance. ............. 122 Figure 13: Performance in Exp. 4 on the direct measures when separating performance by whether or not the participant reported noticing a pattern. (a) 2AFC performance by patterns seen. Percentages indicate the percent of participants noticing or not noticing a pattern. (b) Creation performance by patterns seen. ........................................................................................................................................... 123 1 Chapter 1: Learning from repeated exposure Events in the world are not random. What we see, hear, touch, taste, and smell is replete with regularities that are only evident over time. We can learn these regularities from repeated exposure either gradually, over weeks and potentially hundreds of repetitions, or rapidly, within minutes and only a few repetitions. Though repetition benefits learning over both timescales (Uri Hasson et al., 2015), this thesis focuses on rapid learning, as when we recognize the voice of a person we met earlier at a party or the distinctive visual features unifying a particular movie scene. Repeated exposure updates our internal models of the world and allows us to generate predictions about the future (Clark, 2013; de Lange et al., 2010; Heilbron & Chait, 2017; de Lange et al., 2018; Richter & de Lange, 2019). After repeated exposure, infants and rats can extract abstract relationships that generalize to completely novel stimuli (R. A. Murphy et al., 2008). For example, infants can generalize from a brief exposure to a sound pattern (e.g. “ga ti ga”) to completely different sounds in the same pattern (e.g. "wo fe wo"; Marcus et al., 1999). Repeated experiences enable predictions even with imperfect or incomplete data (e.g. when driving successfully on a foggy day). This phenomenon suggests that repeated experiences can facilitate more accurate decision making, integrating information across multiple sensory modalities to successfully predict the environment. Further, regularity learning can reduce our response times to repeated or expected events (Kalanit Grill-Spector et al., 2006; Bloxom, 1984), allowing us to act more efficiently in the world (e.g. reacting to a traffic light turning green). 2 In this thesis, I investigate when and how people learn from repeated exposure, focusing on the type of learning that has been studied, over the past 25 years, under the umbrella term “statistical learning” (Saffran, Aslin, et al., 1996; Saffran & Kirkham, 2018; Frost et al., 2019). The term “statistical learning” has evolved from referencing the ability to segment spoken words based upon conditional probabilities to, more broadly, the process of extracting regularities from our environment. Despite the conceptual broadening, the experimental paradigm used to study “statistical learning” has remained remarkably consistent. This thesis leverages such a standard paradigm to assess the rapid (approximately 10-minute, 60 repetition) learning of associations within a stream of images. As I elaborate below, statistical learning (SL) has been posited as a ubiquitous and life-long learning process, continually reshaping our representations of the world to change how we parse and remember structured information. In light of the proposed significance of this mechanism, it is important to understand how (if at all) this “statistical” form of learning differs from other forms of learning that have been studied under labels such as “implicit learning” (Cleeremans et al., 1998) and “associative learning” (Thorndike, 1898; Bower & Winzenz, 1970). Since the field of statistical learning has not always employed a single, consistent definition, there is ambiguity about the underlying learning process and how it is distinct from other learning processes. In particular, I present a series of studies to determine whether the visual statistical learning process assessed via the standard paradigm is automatic, implicit, and stimulus-nonspecific. These studies test the conditions under which humans successfully learn about temporal sequence regularities and motivate a reconsideration of how visual statistical learning is conceptualized and studied. 3 A Brief History of Statistical Learning The term “statistical learning” (Saffran, Aslin, et al., 1996) was adopted into human psychological research from the machine learning literature (Saffran & Kirkham, 2018). Infants were theorized to automatically extract statistical regularities from streams of input, with no special attentional or knowledge constraints, via mere exposure (Saffran et al., 1997, 1999b; Aslin, 2017). The original 1996 paper by Saffran and colleagues demonstrated that infants could listen to an unsegmented stream of syllables (e.g. bi-da-ku-pa-do-ti-go-la-bu-bi-da-ku-…) and extract the embedded statistical structure (Saffran, Aslin, et al., 1996). Specifically, four fake “words”, consisting of 3 syllables each (e.g. bi-da-ku), were pseudo-randomly repeated using the same timing between all syllables, regardless of whether a syllable was within a word (e.g. “da”), began a word (e.g. “bi”), or ended a word (e.g. “ku”). After only two minutes of listening, infants reliably expressed a preference for combinations of syllables they had never heard, a sign of familiarization to the originally presented “words”. The work of Saffran, Aslin, Newport, and others had an immediate impact, both in language acquisition theory as well as learning theory in general. Initially, the Saffran findings were taken as evidence for the “empiricist” view of language acquisition in the raging “nativism vs empiricism” debate (Spelke & Newport, 1998; Wagner, 2001). The Saffran findings were interpreted as evidence for an innate learning process, which could explain some of the language-learning capabilities otherwise attributed to an innate language module (Saffran, 2003; Frost et al., 2015; c.f. Chomsky, 1976). More influentially, the infant studies and complementary studies in adults (e.g. Saffran, Newport, et al., 1996) were referenced as powerful evidence for a form of learning 4 that was rapid (occurring in <3 minutes) and incidental (i.e. unintentional and unconscious) and that led to nondeclarative knowledge (Schacter, 1987, 2019; Squire & Zola, 1996; Fig. 1). This constellation of features differentiated statistical learning from the standard forms of associative learning which were thought to be either (a) rapid and intentional, and likely dependent upon medial temporal lobe structures, or (b) slower and unintentional, and likely dependent upon neocortical systems (McClelland et al., 1995). Figure 1: Explicit and Implicit Long-Term Memory (from Squire & Zola, PNAS 1996; Copyright 1996 National Academy of Sciences) Follow-up studies were soon conducted on non-linguistic auditory regularities (Saffran et al., 1999b) and in other modalities, most frequently vision (Olson & Chun, 2001; Kirkham et al., 2002; Fiser & Aslin, 2002). In addition, the field of statistical learning moved from investigating regularities unfolding over time (e.g. in language) with regularities unfolding over visual space (Jiménez & Méndez, 1999; Jiang & Chun, 2001; Fiser & Aslin, 2001, 2005; Fiser et al., 2007; Zhao et al., 2013). The ability to learn regularities with non-linguistic content and across modalities led 5 researchers to conclude that statistical learning was a general learning process that operated across many modalities and stimulus types (Kirkham et al., 2002). In parallel, related lines of research investigated implicit learning of statistical structure using established paradigms such as serial reaction time tasks (SRTTs; Nissen & Bullemer, 1987) and artificial grammar learning (AGL; A. S. Reber, 1967). As statistical learning research has grown in popularity and prominence (Turk-Browne, 2012; Frost et al., 2019), debate has sprung up over many of its foundational assumptions. Is SL automatic (Turk-Browne et al., 2005)? Is the resultant knowledge implicit, explicit, or a mixture of both (P. J. Reber, 2013; Batterink, Reber, Neville, et al., 2015; Bays et al., 2015)? Is SL different from implicit learning at all (Pierre Perruchet & Pacton, 2006)? Is SL a unitary process or a constellation of multiple processes (Arciuli, 2017; Thiessen, 2017; Christopher M. Conway, 2020)? Is SL a domain-general process, is it modality-specific, or can it operate in both manners (Frost et al., 2015)? How does SL ability vary across individuals (Siegelman & Frost, 2015)? Is SL a rapid or gradual process (Turk-Browne, 2012)? Do infants use a different learning process from adults (Thiessen, 2010; Raz & Saxe, 2020)? Are there different neural bases for distinct forms of statistical learning (Karuza et al., 2013; Schapiro et al., 2014; Covington et al., 2018)? How does prior knowledge impact SL (Finn & Hudson Kam, 2008; Emberson & Rubinstein, 2016; Siegelman, Bogaerts, Elazar, et al., 2018)? The variety of debates concerning statistical learning makes evident the lack of a commonly agreed upon definition (Frost et al., 2019). To attempt a broad definition, we suggest that SL is generally held to be a cognitive process (or processes) which enables animals (typically humans) to automatically (i.e. without intent or effort) learn environmental regularities which unfold over 6 time (Saffran et al., 1997; Turk-Browne et al., 2005; Turk-Browne, 2012) without feedback (Aslin, 2017). However, SL can also include learning of spatial (Fiser & Aslin, 2001) or motoric (Christopher M. Conway & Christiansen, 2005) regularities. This definition encompasses a broad swath of learning. It includes any information that is learned after more than one presentation and excludes information that is intentionally studied. Finally, SL is thought to occur across all ages, from infants to seniors (Kirkham et al., 2002; Campbell et al., 2012), and is not differentiated by developmental stage (Gómez, 2002; Raz & Saxe, 2020; c.f. Shufaniya & Arnon, 2018). A fundamental challenge for statistical learning research is that, in the absence of a single commonly agreed upon definition, and with the gradual changes in definitions over time, the underlying processes supporting SL may differ from task to task, even if all of those tasks employ variants of the standard research paradigm. This is problematic, not only because there will not be a single answer to the properties of SL if it is not a coherent entity, but also because it becomes very difficult to situate SL results with respect to the rest of the learning literature. In particular, can we determine which processes are distinctive to SL, and separate from the learning processes described in prior literatures on explicit and implicit human memory (e.g. Ebbinghaus, 1913; Schacter, 1987; Musen & Treisman, 1990; Henke, 2010). Standard Paradigm For Investigating Statistical Learning The standard paradigm for investigating statistical learning follows the format of the original 1996 Saffran study. The standard paradigm has two phases: exposure and test. Before the exposure phase, participants are told they need to attend to a stream of stimuli (e.g. sounds or images) 7 which will be presented one-by-one over several minutes. Then, they are told that they will be asked questions about the stimuli afterwards, so they need to pay attention. The exposure phase consists of a stream of stimuli that contains some regular statistical structure, but participants are not informed about the presence of any structure or regularities in the stream. For example, a stream could consist of hundreds of images, constructed with sequences of 3 images which always occur in the same order, but where several 3-image sequences are presented pseudo-randomly (e.g. A1-A2-A3-B1-B2-B3-C1-C2-C3-A1-A2-A3-C1-C2-C3-…). Importantly, the images are presented unsegmented: the inter-stimulus intervals do not vary within or between the 3-image sequences. After the exposure phase, participants are presented with one or more surprise assessments of learning. Thus, the only way participants perform well on any of the surprise assessments is by learning the embedded statistical structure from the exposure phase, despite being naïve to its existence. One major change from the original paradigm, in addition to testing visual regularities, is the addition of a cover task to the exposure phase. Though participants in the original studies were unable to verbalize the statistical structure after exposure (Saffran et al., 1997), they nonetheless demonstrated knowledge of the regularities via looking times or forced-choice tests of recognition memory. To increase the probability that participants remain naïve to the embedded structure, and to reduce their use of overt learning strategies, cover tasks were introduced to the exposure phase (Turk-Browne et al., 2005). The introduction of a cover task was meant to ensure that the learning process under investigation remained automatic and implicit (Turk-Browne et al., 2005). 8 There are a few standard measures of learning in statistical learning paradigms. Infant research assesses learning using looking times away from familiarized regularities (Saffran, Aslin, et al., 1996; Saffran et al., 1999b; Marcus et al., 1999; Kirkham et al., 2002). On the other hand, research with adults often tests familiarity with forced-choice tasks (Olson & Chun, 2001; Fiser & Aslin, 2002; Turk-Browne et al., 2005; Brady & Oliva, 2008; Kim et al., 2009; Turk-Browne et al., 2009; Arciuli & Simpson, 2012; Bertels et al., 2012; Zhao et al., 2013; Batterink, Reber, & Paller, 2015; Bays et al., 2015; Arciuli et al., 2014; Schapiro et al., 2014; Stevens et al., 2014; Musz et al., 2015; Otsuka et al., 2016; Siegelman, Bogaerts, Kronenfeld, et al., 2018; Vickery et al., 2018). Within forced-choice tasks, the two-alternative forced-choice (2AFC) task is most common. In a 2AFC trial, participants are asked to choose between two sequences of images; one that followed the statistical structure of the exposure phase and one that did not. Since participants have a 50/50 chance of choosing correctly, participants must correctly select the “real” sequence much more than 50% of the time. Infant looking-times are thought to clearly reveal implicit knowledge, as the infant is thought to unintentionally exhibit preferences which reveal learning (Yott & Poulin-Dubois, 2016). However, performance in the 2AFC task, especially in adults, provides ambiguous evidence about the nature of the resultant statistical knowledge. While forced-choice recognition testing has been previously employed as an index of explicit memory (Musen & Treisman, 1990), the test captures some aspects of both implicit knowledge (e.g. inability to verbalize) and explicit knowledge (e.g. making an explicit decision as to the “correct” choice). Turk-Browne and colleagues sum up this ambiguity by stating that the 2AFC task “is an odd sort of dependent measure to use, because it essentially asks observers to make an explicit judgement about implicitly learned relationships” (Turk-Browne et al., 2005). Thus, in hopes of isolating more implicit 9 knowledge, experimenters introduced a more indirect measure of statistical learning, the target detection (TD) task. In a standard TD task, participants are tested with about 90 trials, but the number of trials has ranged from 32 up to 288 (Olson & Chun, 2001; Turk-Browne et al., 2005; Kim et al., 2009; Bertels et al., 2012; Campbell et al., 2012; Barakat et al., 2013; Franco et al., 2015; Batterink, Reber, & Paller, 2015; Bays et al., 2015; Musz et al., 2015; Otsuka et al., 2016; Batterink, 2017; Otsuka & Saiki, 2017). At the start of each trial, participants are cued with one of the stimuli from the exposure phase as the trial-specific “target”. Next, participants self-initiate the trial. Each trial consists of approximately 12-24 stimuli, but implementations range from 7 to 48 elements (Kim et al., 2009; Batterink, 2017). Participants are instructed that, upon detection of the cued target, they should press a key as rapidly as possible. For implementations with longer stream lengths (>15 elements), the target may appear more than once in the same trial. Though response times are what is measured in a TD task, the dependent variable for the TD task is not the raw response times, but rather the facilitation between response times to predictable stimuli relative to less predictable stimuli. For example, “A3” is perfectly predictable if the participant saw A1-A2-A3 many times during exposure, whereas “A1” could follow several different stimuli. Participants can express learning using all three measures described above: infant looking times, forced choice tasks, and target detection tasks. In addition, the knowledge acquired via SL is thought to last at least 24 hours, as measured in both the forced choice (Arciuli & Simpson, 2012) and TD tasks (Kim et al., 2009). Research on statistical learning has primarily used identical or similar paradigms to those described above with surprisingly little deviation (Frost et al., 2019). While this provides an excellent opportunity to compare between studies, and assess the 10 conditions under which humans can learn regularities (Frost et al., 2019), any deficiencies and implicit assumptions in the paradigm have an outsized effect in the theoretical conclusions drawn from the field. Thus, while the standard approach has provided insight into whether statistical learning occurs or not (Conway, 2020; Fig. 2), there has not yet been a systematic investigation into how participants are engaging with the regularities to-be-learned. To delve more deeply into this question, “statistical learning” must be more concretely operationalized. Figure 2: Three possible orthogonal dimensions of statistical learning (Reprinted from Conway, Neuroscience & Biobehavioral Reviews 2020). What are the defining characteristics of statistical learning? The lack of a standard definition for statistical learning was recently highlighted by Frost and colleagues (Frost et al., 2019). They suggested that, under some definitions, almost all varieties of learning would qualify, but they chose to constrain their definition to “perceiving and learning any 11 forms of patterning in the environment that are either spatial or temporal in nature” (Frost et al., 2019). Though this definition by Frost and colleagues certainly accommodates much of the work ongoing in the field, it is still so broad that it remains difficult to relate SL findings with other learning phenomena and their underlying neural processes. How does statistical learning differ from implicit or associative or category learning? Some argue that statistical learning, assessed using serial reaction time tasks, is not much different from the learning studied in classical conditioning (T. Gureckis & Love, 2007). Others argue there is little difference between implicit and statistical learning (Pierre Perruchet & Pacton, 2006). Still others hypothesize that statistical learning is simply a psychological construct for Hebbian learning (Hebb, 1949; P. J. Reber, 2013). In order to understood if and how statistical learning differs from other learning processes we must identify its distinctive features. In the sections that follow, therefore, we review a set of characteristics that could be used to distinguish SL, and we further consider how each of these may be operationalized. We consider three characteristics which (jointly) could distinguish SL from other learning processes: (1) the learning is automatic (i.e. without intent or awareness), (2) the resultant knowledge is implicit (i.e. non-declarative), and (3) the learning is stimulus-nonspecific (i.e. occurs regardless of what stimuli are being associated, across all modalities). Automaticity Many statistical learning researchers indirectly or directly include the concept of “automaticity” in their definitions of statistical learning (Saffran et al., 1999b; Fiser & Aslin, 2002; Turk-Browne et al., 12 2005, 2010; Aslin & Newport, 2012; Batterink, Reber, & Paller, 2015; Batterink et al., 2019; Christopher M. Conway, 2020; Avarguès-Weber et al., 2020). But what makes a learning process “automatic”? Following previous work (Hasher & Zacks, 1979; Turk-Browne et al., 2005; Christopher M. Conway, 2020), an “automatic” process is one that: (a) operates continuously and unintentionally, (b) does not interfere with other processing, and (c) arises unconsciously. These three aspects can be operationalized into the learning being: (1) “always on”, (2) invariant to current processing, and (3) incidental. Despite the prevalence of the assumption of automaticity, the understanding of automaticity has changed over the past two decades. The original conceptualization of automaticity was that the learning occurred via mere exposure, and thus regardless of attention to the stimuli (e.g. Saffran et al., 1997). In the subsequent decade, operating under this assumption of automaticity, the range of cover tasks in statistical learning research proliferated. It was not until the pioneering studies of Turk-Browne et al. (2005) and Toro et al. (2005) that the question was directly addressed (Turk-Browne et al., 2005; Toro et al., 2005). Both studies concluded that SL was not an entirely automatic process, because it required selective attention. At the same time, Turk-Browne et al. (2005) did note that there was a sense in which SL was automatic: because once selective attention was deployed to the stimuli, the learning could occur in parallel with a competing cover task, and its effects could be demonstrated using both indirect and direct tasks. Since these seminal works, the range of cover tasks has continued to grow (e.g. Turk-Browne et al., 2009, 2010; Zhao et al., 2013). While learning remains evident with these new cover tasks, the extent to which the visual statistical learning process is automatic is still debated (e.g. Musz et al., 2015). 13 Perhaps due to the ambiguity around the criteria for automaticity, several research groups to remain undecided about whether the learning is automatic. Reviews by leading researchers mention that statistical learning is typically conceptualized as automatic (e.g Saffran & Kirkham, 2018; Frost et al., 2015, 2019), while other summaries make no claims about automaticity (e.g. Daltrozzo & Conway, 2014; Karuza et al., 2014; Arciuli, 2017; Batterink et al., 2019). In general, statistical learning research has not closely considered how participants cognitively engage with the stimulus stream in the exposure phase of statistical learning studies and how attention interacts with automaticity. Motivated by the uncertainty concerning the automaticity of statistical learning, and the role played by the learner’s attentional state, we examine these questions with a series of high-powered experiments in Chapters 2 and 3. Chapter 2 reports the results of two experiments (Exps. 1 & 2). In Exp. 1, we present data supporting the view that selective attention to the stimuli is necessary, but suggest further that selective attention is insufficient for SL to manifest (c.f. Pacton & Perruchet, 2008). Exp. 1 suggests that the learning process is not “always on”. In Exp. 2, we show that SL is not invariant to current processing, instead the cover task dramatically affects learning. Implicitness Statistical learning researchers agree that statistical learning can occur implicitly, though some do not require this characteristic as a necessary aspect of SL (e.g. Thiessen, 2017; Frost et al., 2019; Christopher M. Conway, 2020). A statement that “statistical learning is implicit” can be interpreted to refer to both the learning process (as incidental, unintentional, and unconscious) and/or the 14 resultant implicit (nondeclarative) memory (e.g. Fiser & Aslin, 2002; Pierre Perruchet & Pacton, 2006; Turk-Browne et al., 2005, 2010; Aslin & Newport, 2012; Arciuli, 2017). However, within a single study, researchers will typically only focus on one of these aspects of statistical learning: whether it is an “implicit learning process” or whether a task resulted in “implicit knowledge” (e.g. Karuza et al., 2014; Aslin, 2017). There is disagreement as to whether and how SL differs from implicit learning and whether the resultant knowledge is always implicit. Most researchers agree that statistical learning occurs incidentally (Fiser & Aslin, 2002; Turk-Browne et al., 2005; Aslin, 2017; Christopher M. Conway, 2020). Going further, Perruchet and Pacton state that both statistical and implicit learning rely upon the same learning processes (Pierre Perruchet & Pacton, 2006). In a similar, but distinct view, Batterink and colleagues argue that statistical learning is a subset of implicit learning (Batterink et al., 2019). The resultant knowledge of the statistical learning process is conceptualized in a similar manner to Figure 1 (Squire & Zola, 1996). “Implicit” knowledge is equivalent to nondeclarative memory, whereas “explicit” knowledge is equivalent to declarative memory. Another view, by Cleeremans, is that the implicit/explicit distinction is moot (at least in terms of the knowledge resulting from learning, if not the learning process itself) because the key part of the definition is about regularity extraction, not about the meta-cognitive access to the learning process or resultant knowledge (Cleeremans, 2006; Daltrozzo & Conway, 2014). Effortful and explicit forms of learning are known to improve with repeated exposure (Ebbinghaus, 1913), and so the distinctive status of statistical learning depends crucially on the claim that it is an incidental process producing implicit knowledge. In turn, the claim that statistical learning produces implicit knowledge rests on data showing that participants exhibit (i) knowledge of 15 embedded statistical structure via indirect measures and (ii) absence of verbalizable knowledge (Turk-Browne et al., 2005; Andrew P. Yonelinas & Jacoby, 2012). Of course, it is possible, and even likely that implicit knowledge co-exists with explicit knowledge in many paradigms, but demonstrating implicit knowledge in such cases is much more difficult, because the results of indirect tests become ambiguous when participants possess explicit knowledge. In Chapter 3 we demonstrate that some of the key data previously taken as evidence of implicit knowledge arising from visual statistical learning, can be explained via confounds in the response time tasks used to assess implicit knowledge. To the extent there is a useful distinction between implicit and explicit statistical learning processes, it appears that effortful learning after instruction can improve auditory statistical learning (Gebhart et al., 2009; Finn et al., 2014), which would suggest that implicit and explicit learning processes interact. The most popular view seems to be that statistical learning is composed of multiple processes, some of which may be explicit (Daltrozzo & Conway, 2014). One apparent difference between how statistical and implicit learning is tested experimentally is the number of stimulus repetitions during the exposure phase. Early work in the field of implicit learning mostly tested effects of one-time stimulus presentation (e.g. priming, savings in relearning; Schacter, 1987, 2019). A few studies, however did investigate learning over repeated sessions or exposures. For example, people suffering from hippocampal amnesia were able to learn sequences of computer programming commands over repeated sessions (e.g. Glisky et al., 1986). Another exception to the rule was a study of neurotypical adults on the effects of repetition on cued (explicit) vs free (less explicit) recall (Schacter & McGlynn, 1989). Participants studied familiar and unfamiliar word pairs (American & British idioms) in one of four possible study tasks. 16 The study tasks varied in either directed attention to the pairing (e.g. reading a sentence where the pair was defined) or a more indirect exposure (e.g. counting the number of vowels). The authors found that repetition (between 2, 4, or 8) monotonically improved performance in both the cued and free recall. Similarly, reading repeated, paired words results in better memory than unrepeated paired words, with especially improved memory when the words require some additional processing (e.g. when exchanging some letters for blanks; Cuddy & Jacoby, 1982). While the results of these experiments may reflect implicit memory, it is possible that, just like statistical learning, there is a mixture of implicit and explicit learning processes and resultant knowledge. Many early studies of “implicit memory” lacked careful debriefing on the extent of explicit knowledge they possessed. For example, in many of Schachter’s seminal studies, participants were told that their memory would be tested after exposure, but the experiments did not assess how many participants employed an explicit strategy. There are two further experimental paradigms intended to investigate implicit learning that involve stimulus repetition and pre-date statistical learning research: artificial grammar learning (AGL; Fig. 3; A. S. Reber, 1967) and serial reaction time (SRT) tasks (Nissen & Bullemer, 1987). How does SL differ from the learning processes demonstrated in these tasks? Some differences between the standard SL paradigm and standard AGL/SRT tasks are summarized in Table 1 (Christopher M. Conway, 2020; Karuza et al., 2014). Conway further notes that motor and perceptual learning have different time courses and that motor action can lead to an improved rate and amount of learning. This contrasts with the standard statistical learning research paradigm which has no motoric component. 17 Table 1: Paradigms for studying implicit (and statistical) learning. Modified from Conway (2020). Standard Research Paradigm Artificial Grammar Learning (AGL) Serial Reaction Time (SRT) Dependent Variable Forced Choice, RT Facilitation Explicit judgement of grammaticality RT, Response accuracy Modality Perceptual (any) & Cross-modal Perceptual (any) Visual-motor (auditory has been used successfully) Input Structure Transition Probabilities Artificial grammar Repeating sequences Generalization at Test Mixed Mixed No In the following subsections, I briefly highlight two of the standard research paradigms for implicit learning (artificial grammars and Serial Reaction Time tasks) to clarify how they differ from the standard research paradigm in statistical learning. Generally, paradigms such as artificial grammar learning (A. S. Reber, 1967) and the serial reaction time task (Nissen & Bullemer, 1987) require participants to continually generate actions or decisions that follow the same regularities that are later tested (as noted in Olson & Chun, 2001), versus passive exposure in the standard SL paradigm. Artificial Grammar Learning Artificial grammar learning (AGL) paradigms are used to study the implicit acquisition of rules dictating transitions between sequence elements (Fig. 3; A. S. Reber, 1967). For example, all sequences in Fig. 3 begin at state S0 and the transitions between further states are associated with the addition of a symbol to the generated sequence (labeled T, V, X, P and S). Hence, each sequence begins with either T or V. An example of a valid sequence is “TPPPTS”. The sequence begins with T, as the system transitions from S0 to S1, and then three P’s are inserted as the system 18 transitions recurrently from S1 back to S1. The next state is S2, so another T is appended during this transition. Finally, this sequence concludes with the symbol S at S0’. Participants in artificial grammar learning studies are typically unable to explain the rules subsequent to the exposure phase. This property easily places this type of learning on the “implicit” side of Fig. 1. However, Arthur Reber argues that since subjects were not given instructions on how to memorize the sequences, they likely created their own “organizational schemes” to remember the sequences they were trained upon. Since the initial studies (Braine, 1963; A. S. Reber, 1967), researchers have noted that artificial grammar learning is unlikely to be entirely implicit, potentially requiring a combination of procedural and declarative memory (Christopher M. Conway, 2020). Insofar as the sequence grammaticality is generated and judged motorically, the learning mechanisms may be motor-specific and implicit (Christopher M. Conway & Christiansen, 2005; C. M. Conway & Christiansen, 2006). When the AGL paradigm involves providing feedback to participants or as participants attempt to organize the sequences using mnemonics, learning may be more explicit. Figure 3: Artificial grammar learning paradigm (Reprinted from A. S. Reber, J. Verbal Learning & Verbal Behavior 1967) 19 Serial Reaction Time Tasks Prior to the emergence of the statistical learning literature, Serial Reaction Time (SRT) tasks were a popular way to measure implicit sequence learning. As in AGL, participants are typically unable to verbalize the learning, but the SRT tasks have a more obviously motoric and procedural learning character. In the original SRT study, participants were seated in front of an array of four lights arranged horizontally, with four matched buttons (one for each light; Nissen & Bullemer, 1987). Participants were split into two conditions (“random sequence” and “ordered sequence”). In the ordered sequence, participants saw a repeating 10-element light sequence. In the “random” sequence, participants saw several, unstructured 10-element light sequences. Participants were instructed to press the button matched with a light, as quickly as possible, whenever a light came on. The dependent variable was the reaction time of pressing the button. Participants in the “ordered” condition demonstrated faster reaction times relative to the “random” condition. Motoric aspects of the original SRT paradigm can contribute to the learning (Song et al., 2008). In addition, explicit knowledge can arise during an SRT task and can be facilitated with cues (Song et al., 2007). However, Song and colleagues (2008) demonstrated that simply viewing a sequence is sufficient to produce reaction time speeding, equivalent to a motoric condition, and without explicit knowledge arising. Despite the apparent indifference of simply observing sequences and motorically mimicking them, Song and colleagues do not rule out the possibility of rapid motoric learning during test for the participants in the “observe” condition. 20 Stimulus Non-specificity Almost all statistical learning researchers agree that it can operate across stimulus types and sensory modalities (Frost et al., 2019; Christopher M. Conway, 2020). Statistical learning is distinguished by the broadness of the class of stimuli over which it is supposed to operate: it has been claimed to operate across spatial regularities (Fiser & Aslin, 2001), linguistic regularities (Saffran, Aslin, et al., 1996), tones (Saffran et al., 1999b), music (Peretz et al., 2012), somatosensation (Christopher M. Conway & Christiansen, 2005), images (Fiser & Aslin, 2002), and across almost any sensory input tested. This breadth of learning examples, suggests that SL automatically binds repeated arbitrary associations, regardless of stimulus presented. Modality-specific or domain general? An important question in the field of statistical learning is whether the learning process is mechanistically equivalent across stimuli or only functionally equivalent. In other words, to what extent do the neural mechanisms overlap for processing stimuli from different domains? Consider that the “statistical learning mechanism can operate on both linguistic and non-linguistic stimuli” (Saffran et al., 1999b). This could suggest that the same neural mechanisms operate over all auditory stimuli equivalently. Similarly, statistical learning effects are observed for stimuli presented in the auditory, visual, and somatosensory domains (Saffran, Aslin, et al., 1996; Baker et al., 2004; Christopher M. Conway & Christiansen, 2005). Does this breadth of domains suggest that SL operates identically (and centrally) across all input? 21 Views on modality specificity have changed over the decades. In early studies of implicit memory, the theory was that implicit learning occurred in modality-specific (not domain general) ways (e.g. Schacter & Graf, 1989). When the field of statistical learning began, the earliest studies made no claims as to the multi-component nature of SL, though some suggested a domain general, unitary process (e.g. Kirkham et al., 2002). Similarly, a few subsequent SL theories suggested a seemingly unitary and domain general process (e.g. Pierre Perruchet & Pacton, 2006; Aslin & Newport, 2012). By the mid-2010’s however, researchers began to coalesce around the view that statistical learning was actually a family of processes (Daltrozzo & Conway, 2014; Bays et al., 2015; Thiessen, 2017; Arciuli, 2017). This belief comports well with evidence from related fields like working memory and judgements of temporal order, which have shown modality-specific effects (Gmeindl et al., 2011), but are known to permit multi-modal processing (Baddeley, 2012). The lack of a clear definition for statistical learning affects the ability to address the degree of overlap in neural mechanisms. When statistical learning is defined as “all phenomena related to perceiving and learning any forms of patterning in the environment that are either spatial or temporal in nature” (Frost et al., 2019), then it would seem to be highly domain general. However, under the functional definition, SL is a nebulous family of processes that operates on any input and may only exclude single shot learning. If so, then it could be supported by a very general class of mechanism that shape and re-shape information processing over time, broadly analogous to (or even identical to) the broadest notion of cortical plasticity (e.g. P. J. Reber, 2013). This question of whether statistical learning is a unitary, domain-general process or a cognitive descriptor of a family of processes operating in each modality separately is still unsettled. Further, it is possible that there exist multiple domain-general processes, so there is not a single dimension 22 between theories of whether statistical learning is a unitary or multi-component process. The current view is that statistical learning is not a single cognitive process which operates identically on any input, but does not specify whether the family of learning processes are specific to modality or are a combination of modality-specific and input-insensitive processes. The current evidence suggests the latter (Frost et al., 2015; Christopher M. Conway, 2020). What is not statistical learning? Many types of learning have been designated as “statistical learning”, and so it is important to specify the boundary conditions under which certain types of learning are excluded from the umbrella term. Two clear examples are motor learning and reinforcement learning. Although SL processes may be involved in SRT and AGL tasks, these tasks often require motoric actions that are associated with the statistics to-be-learned. If statistical learning is meant to be understood as a process distinct from motor learning (e.g. Körding & Wolpert, 2004), then it should be studied using tasks which do not associate motor activity with the regularities to-be-learned. It is unsurprising that procedural memory is elicited after multiple repetitions and results in implicit knowledge. Similarly, most definitions of SL require the learning to occur “without feedback” (Aslin, 2017), so studies in reinforcement learning and classical conditioning are excluded (e.g. Schultz et al., 1997). Of note, some reward learning tasks have been shown to share neural processes with SL, despite being conceptually separable (Shohamy & Turk-Browne, 2013). Turk-Browne and colleagues (2010) distinguished statistical learning from other forms of learning as follows, “Statistical learning differs from other forms of learning because regularities are not 23 presegmented into discrete trials (cf. paired associate learning), the underlying structure is stimulus specific (cf. artificial grammar learning), and the learning is automatic and incidental (cf. some varieties of category learning). These features allow statistical learning to operate over naturalistic sensory input without engaging deliberate/ conscious effort”. Although these delineations accurately reflect common beliefs in the field, they lay down boundaries which may be too exclusive (e.g. excluding paradigms which test on segmented trials or warn participants of upcoming structure) or inclusive (e.g. do the stimuli need to be novel?, must the learning be unrewarded?). Participants performing SL tasks are supposed to acquire knowledge about regularities without the intent to learn. Therefore, some early tasks (whose instructions warned participants of upcoming patterns) may no longer be considered valid examples of statistical learning (Saffran, Newport, et al., 1996; McNealy et al., 2006). Additionally, these design decisions render it nearly impossible to demonstrate that the subsequent learning was automatic or incidental. As we will argue in Chapter 2, not all cover tasks are equal: instead, learning appears to be modulated by the extent to which the instructions and cover task demand attention to temporal structure. We will further argue that, when exposure paradigms involve continuous measurements or judgments that are correlated with the regularities-to-be-learned, this will boost statistical learning. Not all experimental paradigms which possess a statistical structure, qualify as a statistical learning paradigm under the definitions given above. For example, oddball effects in auditory tone sequences imply that some level of statistical regularities are being recognized and retained in our brains, but this phenomenon occupies a space outside of traditional statistical learning (R. Näätänen et al., 1978). The oddball paradigm typically consists of two tones (e.g. 1000Hz and 24 1100Hz). One tone is played repeatedly (i.e. the standard) and then the other tone is played (i.e. the deviant). When the deviant tone is played, a neural signal is observed which seems to represent a prediction violation (R. Näätänen et al., 2007; c.f. May & Tiitinen, 2010). An oddball response can be generated extremely rapidly (within a few trials) and are therefore considered to be automatic; that is, that the “deviation” response is thought to occur without intent. One of the main reasons it is not an obvious candidate for SL is because the oddball is so salient that it almost exclusively results in explicit knowledge of the oddball (e.g. as assessed with the P300 ERP component). Additionally, it is not thought to be an associative learning process so much as a “deviant detection” process. Other forms of automatic, but not “statistical” learning can be found in stimulus-specific adaptation to repetitive stimuli (e.g. Vinken et al., 2017). Similarly, predictive processing theories suggest that learning occurs through repeated exposure to stimuli (Todorovic et al., 2011), but is generally considered to be broader than “statistical learning” because of top-down influences. How does statistical learning differ from associative learning? As stated earlier, Turk-Browne and colleagues attempt to separate statistical learning and associative learning primarily on the basis of experimental paradigm. However, they raise the possibility that the same neural structures are responsible for all associative learning, regardless of whether the learning occurs implicitly or explicitly (Turk-Browne et al., 2010). Indeed, the fact that knowledge acquired during a task is accessible explicitly does not determine the location of the memory “store”; explicitly accessible and inaccessible knowledge could rely on a common store, which could include medial temporal lobe structures. 25 Neural underpinnings of statistical learning Due to the broad, overlapping, and sometimes conflicting definitions, it is difficult to determine which neural mechanisms are relevant to and necessary for statistical learning. Additionally, the cartoon distinctions between “implicit” and “explicit” memory in Fig. 1 have led some psychologists to conflate psychological distinctions in learning and their underlying neurophysiological “memory stores”. This conflation has driven broad interest in the claim of psychological implicit learning with a putative cortical learning mechanism. Figure 1 neatly separates implicit and explicit learning into distinct long-term memory structures (c.f. Reder et al., 2009). According to these distinctions, memory for implicitly learned material tends to be located outside the neocortex. However, there are at least two examples of implicit learning which might be neocortically-based. One is priming, where previously presented information facilitates subsequent processing through distinct mechanisms from explicit memory (Schacter, 2019). Priming can occur with multiple types of input, including syntactic priming (where the structure of a sentence facilitates the processing of a syntactically-similar sentence) and semantic priming (where a concept facilitates the processing of a related concept). These phenomena are thought to reflect the activation of an item’s cortical representation (e.g. a particular grammar or concept) and this activation overlaps with the representation of a subsequently facilitated item (Kutas & Federmeier, 2011). Since the activation of a subsequent item is “for free”, there is a reduced need for the otherwise complete cortical activation, and the facilitation can be implicit and endogenously facilitated (Goldrick et al., 2010). A second form of neocortically-based implicit learning (not shown in Fig. 1) is artificial grammar learning (Squire & Zola, 1996). 26 Several important differences have been proposed between cortically- and hippocampally-based learning. While forms of implicit learning, like AGL, might be cortically-based, associative learning is typically considered to be hippocampally-dependent (Degonda et al., 2005). Cortical learning processes are often thought to occur implicitly (i.e. unconsciously), whereas learning dependent upon the hippocampus and medial temporal lobe (MTL) is traditionally associated with explicit knowledge (Fig. 1). Although “incidental” learning is thought to be unintentional and implicit, it can also be MTL-dependent (Stark & Okado, 2003), suggesting that the learning process could rely upon the same neural substrate as more “explicit” learning. Traditional learning theory suggested that the cortex learns stimulus associations slowly, while the hippocampus learns rapidly (McClelland et al., 1995). However, recent findings have suggested that, in some situations, the cortex can rapidly associate stimuli (Coutanche & Thompson-Schill, 2014; c.f. Cooper et al., 2019). Since most SL paradigms are ostensibly implicit and require the rapid learning of associations, it remains unclear whether the learning is hippocampally-dependent, cortically-dependent, or a mixture of the two. Do potentially multiple statistical learning processes share underlying neural processes? For example, the overlap between temporal and spatial statistical learning is unclear. Opinions diverge as to whether statistical learning operates over both dimensions equivalently. Both may be dependent upon the hippocampus, though the hippocampus has long been implicated in spatial activities (O’Keefe & Nadel, 1978). Recent work has suggested that temporal regularity learning is also hippocampally dependent (Turk-Browne & Scholl, 2009; Schapiro et al., 2014, 2016). Although this thesis reports only behavioral data, its theoretical focus on automatic learning and implicit knowledge can influence our interpretation of the neural underpinnings of SL (Schapiro 27 et al., 2012, 2013, 2014; Gavornik & Bear, 2014; Himberger et al., 2018; Richter & de Lange, 2019; Vergnieux & Vogels, 2020). It is tempting to wonder if SL is supported by a hippocampally-based learning mechanism that is automatic, implicit, stimulus-nonspecific? Evidence against this simple explanation can be found in studies with macaques and mice showing that SL can give rise to changes in early visual cortex (Gavornik & Bear, 2014; Vergnieux & Vogels, 2020) and that SL is not always hippocampally-dependent (Covington et al., 2018). Going forward, it seems important to specify subtypes of statistical learning, and to test whether these subtypes are supported by distinct neural processes. There is evidence that many cortical and subcortical regions are involved in statistical learning (Finn et al., 2018). There are clear differences between “associative” learning tasks and “statistical” learning tasks, but it is unclear whether this means they tap distinct learning processes. Associative learning is typically defined in relation to the experimental paradigm (i.e. one-shot learning) or neural region (the medial temporal lobes; Mayes et al., 2007); little consideration is given to repeated stimulus presentation. Specifically, associative learning is typically limited to a simple association, which is consciously perceived, and the learning is either categorized as resulting in semantic or episodic memory. Since associative learning can occur with as little as one pairing between two items, we may naturally think of associative learning processes as more rapid, and statistical learning processes as more gradual. However, it is not clear whether drawing a terminological distinction based on the most common learning tasks (i.e. how we assign names to experimental paradigms) is meaningfully distinguishing the underlying learning processes. Recent research on prior knowledge and schema formation has begun to focus less on one-shot learning and more on how experience modulates information processing (van Kesteren et al., 28 2012; Gilboa & Marlatte, 2017), but the field of statistical learning remains mostly siloed from the study of how repeated exposure to stimuli affects learning (c.f. Brady & Oliva, 2008; Emberson & Rubinstein, 2016). Also relevant is that the associative learning benefits of prior knowledge are explicitly excluded from statistical learning paradigms (Turk-Browne, 2012); the goal is to keep the stimulus choice novel to participants in order to study “pure” statistical learning. Prior Knowledge Some statistical learning studies have shown that the knowledge participants bring into the experiment modulates what they learn (e.g. Emberson & Rubinstein, 2016). In this study, participants saw example images from a few categories (e.g. dogs, fishes, birds, flowers). Using a standard SL paradigm, they paired exemplars from two categories (e.g. dogs-fishes) and then assessed learning. They found that when they paired typical exemplars from the categories, participants learned the pairings significantly better than when atypical exemplars were used. Not only were the participants in the typical exemplar condition more accurate in a forced-choice task than those in the atypical condition, but they also had more explicit knowledge of the category pairings. Prior knowledge can also be important in auditory learning tasks. When learning a spoken language, words that are difficult to pronounce (i.e. different from your native language) are acquired more slowly than words that are more similar to your native language (N. C. Ellis & Sinclair, Susan G., 1996). In working memory tasks, it’s easier to remember non-words that have high “wordlikeness” to your native language relative to low “wordlikeness” (Gathercole et al., 29 1991). These examples highlight how associations between stimuli are influenced by the number of times they are experienced in our environment. Preview of our findings The overarching empirical question of this thesis is: “What are the conditions under which statistical learning is impaired or facilitated?” By leveraging the standard statistical learning research paradigm across multiple experiments, we demonstrate how learning is affected by: the amount of exposure to regularities in a simple perceptual cover task (Experiment 1), the choice of cover task during exposure to regularities (Experiment 2), the sequences used in response time measures of implicit knowledge (Experiment 3), and the choice of stimuli to be associated (Experiment 4). In this series of large-sample experiments, we found that statistical learning was not automatic, even for attended stimuli, and that it was profoundly affected by whether participants attend to (or seek to discover) the temporal aspects of the visual world. Our data suggest a reconsideration of the automaticity of visual regularity learning, and its relationship to processes traditionally studied in the domains of selective attention, relational processing, and basic cortical plasticity (P. J. Reber, 2013). We found that, although regularities in the world around us may be learned without an explicit objective, the learning rate is modulated by many of the same features of attention, engagement, and prior knowledge that are known to shape effortful and directed learning (de Leeuw, 2016). We find that not only is there a potential confound in the previously used methodology for 30 assessing implicit learning, but that the nature of engagement with regularities (which often differs between research groups) has a significant effect on the learning outcomes. Collectively, these data suggest that statistical learning, as measured in standard visual statistical learning paradigms, is not best conceptualized as a single distinctive form of learning. Instead, the learning demonstrated in these studies appears to rely on multiple learning mechanisms, including associative learning and procedural learning, which preceded the conception of statistical learning. Overall, this thesis provides evidence against the claim that the visual statistical learning measured in standard paradigms is an automatic and stimulus-nonspecific learning process giving rise to implicit knowledge. 31 Chapter 2: Visual statistical learning did not appear to be automatic and was strongly modulated by task Portions of this chapter were posted on PsyArXiv (Himberger et al., 2019). Can humans learn “automatically”? It has been argued that not only is statistical learning “automatic”, but that this automaticity is one of the defining characteristics of visual statistical learning (Turk-Browne et al., 2005). The claim is that subjects will passively and even unconsciously extract statistical regularities from streams of input as long as they attend to the stimuli (Toro et al., 2005). Such automatic learning is thought to occur without interfering with other cognitive processing (Hasher & Zacks, 1979). Thus statistical learning is thought to occur “automatically”; that regularities are gradually extracted from a repeated stimulus presentation, without intent or awareness, as long as the stimuli are attended via “mere exposure” (Aslin, 2017). Since the regularities are thought to be “soaked up” by “mere exposure” (Aslin, 2017) and “without trying” (Plante & Gómez, 2018), participants are often asked simply to attend to the stimuli, or to perform an incidental cover task that is unrelated to the sequence structure. Attending to the stimuli is necessary for visual (Turk-Browne et al., 2005; Toro et al., 2005) and auditory statistical learning (Fernandes et al., 2010) to occur in adults, but it remains unclear whether visual statistical learning (VSL) is automatic when attention is passive (i.e. when participants attend to the stimulus features, but do not have to deeply process the stimuli or any relationships between stimuli). In other words, is passive attention sufficient to support VSL? Or is some more active form of processing necessary for learning? 32 Under the assumption of automaticity, we set out to study the effects of repetition on learning (Experiment 1). Our hypothesis was that repeated exposure would benefit learning, as the automatic learning process would have more time to operate. This hypothesis is consistent with previous findings, showing that increasing the number of repetitions during exposure (Fiser & Aslin, 2002) or elongating the stimulus duration (Turk-Browne et al., 2005) leads to better performance on direct measures of visual statistical learning. In contrast to our pre-registered hypotheses and previous reports, however, participants who were exposed to regularities while attending to transient motion in a stream of images (Turk-Browne et al., 2009) exhibited weak evidence of learning using two direct measures (two-alternative forced choice of sequence familiarity and explicit creation of the regularities). Thus, we did not observe (a) evidence of automatic learning, nor (b) evidence of improved learning with more exposure. Given our results from Experiment 1, we hypothesized that visual statistical learning is not entirely automatic, and that, beyond simply attending to the stimuli, visual statistical learning depends substantially on participants’ task goals, specifically in how they process the relationships between stimuli (Vickery et al., 2018). Thus, in Experiment 2, we investigated another aspect of automatic learning, that it occurs regardless of other cognitive processing. We measured visual statistical learning performance across six different task conditions in which participants viewed the same groups of images from Experiment 1. As task instructions across conditions increasingly required participants to attend to the relationships between stimuli, the participants’ performance increased from chance (in Experiment 1) to robust levels of learning. These data suggest that little visual statistical learning occurs unless participants, by necessity or inclination, associate the items 33 in a visual sequence. We conclude that the learning observed in visual statistical learning paradigms is, for the most part, not automatic. Experiment 1: Repetition does not improve learning when task is unrelated to temporal processing Repeated practice and exposure tends to benefit learning (Ericsson et al., 1993; Batterink, 2017). Stronger visual statistical learning effects have been shown after more repetitions (Fiser & Aslin, 2002) or elongated stimulus presentation duration (Turk-Browne et al., 2005). Participants who performed an “implicit” task, counting letters in words that were shown 2, 4, or 8 times, demonstrated increasingly better cued recall performance with repetition amount (Schacter & McGlynn, 1989). “Practice makes perfect” reflects the benefits of multiple repetitions. Here we investigated whether visual statistical learning in adults benefits from repetition. By manipulating the number of repetitions, we investigated how the indirect and direct measures of regularity learning are affected by either: no exposure (0 repetitions), a moderate amount of exposure (60 repetitions), or prolonged exposure (96 repetitions). Consistent with prior studies (Batterink, Reber, Neville, et al., 2015; Batterink, Reber, & Paller, 2015; Bays et al., 2015), we reasoned that if VSL depends on a singular underlying mechanism or process (Kirkham et al., 2002; Perruchet & Pacton, 2006; c.f. Frost, Armstrong, Siegelman, & Christiansen, 2015), then learning outcomes should be correlated across distinct measurements within the same participant. However, if statistical learning draws on multiple mechanisms, then a participant might express learning under one measure, but not under another measure. We used one indirect 34 (Olson & Chun, 2001; Turk-Browne et al., 2005; Kim et al., 2009; Bertels et al., 2012; Batterink, 2017) and two direct (Saffran, Newport, et al., 1996; Fiser & Aslin, 2002; Turk-Browne et al., 2005; Schapiro et al., 2012; Otsuka et al., 2016) dependent measures to assess the learning of visual regularities within each subject: a target detection (TD) task, a two-alternative forced choice (2AFC) task, and a triplet creation (Creation) task (Fig. 4b). As shown previously in visual and auditory paradigms (Batterink, Reber, Neville, et al., 2015; Bays et al., 2015), dissociable learning processes can be teased apart with the careful selection of post-exposure measures such as these. We hypothesized that, (1) the amount of repetition would not modulate learning as assessed by response time speeding, but (2) repetition would benefit performance in both the 2AFC and Creation tasks. As hypothesized, we found that repetition did not modulate learning in response time speeding. In contrast to our second hypothesis, however, we did not observe a benefit of repetition on performance in the two direct measures of visual statistical learning. Our experimental design, hypotheses, exclusion criteria, and analyses were pre-registered using the Open Science Framework (https://osf.io/vtmpb). Method Participants. 360 participants were analyzed, with 120 participants in each of three conditions. As documented in our pre-registration, we performed a power analysis using pilot data and estimated our power at greater than 99% when collecting 120 participants in each condition. Each participant was assigned to either the “0 Repetitions”, “60 Repetitions”, or “96 Repetitions” condition. Exposure amount was defined by how often each triplet (see “Stimuli” below) was 35 shown in the “Initial Exposure” phase: 0, 60 or 96 times. Participants were recruited to participate via an advertisement on Amazon’s Mechanical Turk platform, with the experiment coordinated via psiTurk 2.1.2 (T. M. Gureckis et al., 2016). More detailed recruitment, payment, and demographics are shown in the Supplemental Methods (Supp. Table 1; Supp. Fig. 1). Exclusion Criteria. Online participants require greater compliance and comprehension testing than in-lab participants (Kraut et al., 2004). 602 subjects participated, with 157 subjects excluded and 85 held out to ensure that we had counterbalanced datasets in each condition. Using pre-registered exclusion criteria (Supp. Table 1), we excluded participants for: incorrectly recorded data (n=1), not performing the cover task with at least 50% accuracy (n=23), not performing the Target Detection task with at least 50% accuracy (n=10), changing windows away from the experiment >20 times (n=22), indicating not all stimuli were seen (n=7), indicating previous participation in a similar experiment (n=13), excessive keypresses (e.g. pressed the space bar repeatedly) (n=15), not making a selection in at least 95% of 2AFC test trials (n=21), indicating at least one set of instructions was not understood (n=27), not following instructions (n=15), or taking an excessive amount of time to complete the experiment (n=3). Stimuli. Within each condition, each participant saw a “stimulus set” composed of twelve unique images (Fig. 4a; Schapiro, Kustner, & Turk-Browne, 2012). Images were chosen to ensure no previous associations with the stimuli. The 12 images were organized into 4 sets of three images (i.e. triplets). Thus, we can label the triplets in a stimulus set as A, B, C and D, with each triplet composed of three “fractal” images: (A1, A2, A3); (B1, B2, B3); (C1, C2, C3); (D1, D2, D3). Each participant within a condition saw a different set of stimuli, but stimulus sets were shared across conditions. 36 We counterbalanced the stimulus sets to ensure that each image appeared equally often as the first, second, or third position of a triplet. Thus, each “A” triplet (A1, A2, A3) had two corresponding triplets A’ (A2, A3, A1) and A’’ (A3, A1, A2,) that two other participants in that experimental condition saw. This counterbalancing ensured that triplet position was uncorrelated with image identity across our dataset. Procedure Each participant completed the phases of the experiment (Fig. 4b; Supp. Table 2) in the same order: Pre-experiment Instructions, Initial Exposure, Target Detection (TD) Task, Two-Alternative Forced Choice Task (2AFC), Triplet Creation Task, and Post-experiment Questionnaire. Each phase is described in more detail below. 37 Figure 4: Stimuli & experimental paradigm. (a) The 12 “fractal” images used. (b) Each column shows one phase of the experiment. The "Task" row indicates the participant’s goal in each phase. The "DV" row indicates the primary dependent variable measured in that phase. The experiment was designed so that more direct (more explicit) measures of VSL occurred later. Participants in the “0 Repetitions” condition skipped the “Initial Exposure” phase. Pre-experiment Instructions. After providing informed consent, participants read general instructions regarding the experiment and its expected duration. Participants were instructed that they should dedicate complete attention to the experiment and to attend to each image. Participants were not told anything about regularities. Attend Jiggle Practice: Participants were instructed to watch a stream of images and to detect any image movement from side-to-side (i.e. “jiggling”; Turk-Browne et al., 2009). Participants were instructed to press the space bar within 600ms of a jiggle occurring. Before exposure to the 38 experiment stimuli, participants practiced detecting jiggles in two consecutive stages. In the first stage, participants saw the word “PRACTICE” displayed 20 times, using the same timing as the main exposure phase (i.e. 800ms on, 200ms off). On 4 of the 20 instances, the word jiggled and participants were provided feedback on whether or not they responded to the jiggle quickly enough. The second stage was identical, but no feedback was provided. This second stage was repeated until all four practice jiggles were caught. After successful completion of the practice phase, participants proceeded to the Jiggle Task. Initial Exposure. Participants (except those in the “0 Repetitions” condition) viewed each of the four triplets multiple (60 or 96) times while performing a “Jiggle Task” in which they were asked to detect image motion (Turk-Browne et al., 2009). Participants in the “0 Repetitions” condition proceeded directly from the pre-experiment instructions to the Target Detection task, skipping the initial exposure. To complete the “Jiggle Task”, participants were instructed to detect any transient horizontal motion of an image within a stream of images (i.e. to “catch jiggles”). Participants were instructed to press the space bar whenever a jiggle occurred. The purpose of the Jiggle Task (Fig. 4b, leftmost column) was to diminish the chances that participants would employ explicit strategies to examine the statistical patterns embedded in the stream of images, while ensuring the stimuli were attended. During the Jiggle Task, participants viewed a series of polychromatic, abstract images (Fig. 4a) presented one-by-one. Images were shown for 800ms with an inter-stimulus interval of 200ms. 39 Jiggles began 300ms after stimulus onset with an immediate onset of slight rightward, then leftward, then rightward motion at constant speed. The duration of image motion took 200ms. Jiggles could not occur within 6 images of one another. The exposure phase never began or ended with an image that jiggled. Each image was shown 60 or 96 times (depending upon condition), for a stream of 720 or 1152 images, respectively. The number of jiggles in each condition was equal to the number of repetitions per image (i.e. 60 or 96), thus each of the twelve images was selected to jiggle 1/12th of the time (e.g. in the “60 Repetitions” condition, each of the twelve images jiggled 5 times, for 60 jiggles total). The number of items between consecutive jiggles varied pseudo-randomly, ranging from 6 to 31. Measures of Visual Statistical Learning. After experiencing either 0, 60, or 96 repetitions of each triplet during the initial exposure, participants began the test phase. The order of the three measures (Target Detection, 2AFC, and Creation) was chosen to situate the most indirect measure first, with increasingly direct measures following. If participants were exposed to direct measures of learning first, then their performance in the indirect measure could be influenced by explicit knowledge of the structure in the experiment. Target Detection (TD) Task. The target detection (TD) task (Fig. 4b, second column; Fig. 8a) tested participant response speed to targets embedded within streams of 7-11 images. Stream length was limited to ensure that no images repeated during a trial, avoiding any potential priming effects. These streams contained the same triplet sequences seen during the exposure phase, so that target detection speed might be affected by what participants learned about the sequences. 40 Before the start of each TD trial, a single, trial-specific “target” (one of the 12 unique images shown during the exposure phase) was presented to the participant. Participants initiated each trial by pressing the enter key, then pressing the space bar upon detection of the trial-specific target. Only one target was shown on each trial and each of the 12 images served as a target 6 times. Therefore, participants completed 72 TD trials. Following previous work (Turk-Browne et al., 2005; Musz et al., 2015), stimulus duration and inter-stimulus interval (ISI) were decreased during test (200ms & 40ms, respectively). Participants were asked to respond quickly and accurately. No practice sessions were performed to reduce interference from the introduction of new stimuli. The position of each target within each 7-11 image stream was balanced across trials. Following previous work, the target image never appeared in the first or last three positions of a trial (Turk-Browne et al., 2005). Targets could only appear in stream positions 4-8 and therefore each stream length was 7-11 images. TD Trial Counterbalancing. Triplet positions were balanced by stimulus set (see “Stimuli” above), so images occurred equally in each triplet position. In the TD task, however, we wanted to avoid any potential response times effects of triplet position and where the target appeared in a trial (i.e. stream position). Therefore, we equated target presentation across stream positions. To ensure that any stimulus could appear at any stream position, it was necessary that streams did not always begin with an image that was the first item in a triplet. Given the constraint that trials were between 7-11 images, the stream position of the target dictated (i) the triplet position of the first item and (ii) the total stream length. 41 For further clarity on the process, consider if the trial-specific target is C2 (i.e. the second element in the “C” triplet) and it was to appear in the sixth stream position. This would dictate the trial beginning with an image that was the third position of another triplet (e.g. A3, B1, B2, B3, C1, C2, C3, D1, D2). As no trial had a target in the final three stream positions, the total stream length would be 9 (target position plus three; Fig. 8a, “C2 Length 9”). Each image served as the target twice in position 4, and once in all other stream positions (5, 6, 7 or 8). By making targets appear more often in stream position 4 than in other positions, we aimed to mitigate the RT speeding effect that arises from having a non-constant hazard rate (i.e. having an increasing probability that the target would appear as the next item, given that it had not yet appeared in the stream) and increased expectancy (Perruchet et al., 2006). Collectively, this procedure ensured that all target images (regardless of whether they were in the first, second or third position of their triplet) appeared equally often at each stream position and that there was no relationship between triplet and stream position. The dependent variable was the response time, which we used to calculate response time facilitation between the first and third triplet positions (see “TD Analysis” below). Ensuring stream position is not dictated by triplet position is uncommon (Turk-Browne et al., 2005; Kim et al., 2009; Turk-Browne et al., 2010; Campbell et al., 2012; Bays et al., 2015; Musz et al., 2015) and we investigate our design choices further in Experiment 3. Awareness Probe. Following the TD task, participants were asked if they “noticed any patterns in the stream of images”. If they responded positively, they were asked to “briefly describe” the patterns. Regardless of response, participants then proceeded to the more direct measures of 42 learning. Two-Alternative Forced Choice (2AFC) Task. The 2AFC phase began by informing participants (i) that the images they had seen during the exposure phase had followed a pattern and (ii) that they would need to make judgements about whether groups of images followed that pattern. Participants were told they would choose between two sequences of three images and to choose the group that seemed more familiar. Participants began by practicing the 2AFC task. They were told to use the left or right arrow key to indicate which of two 3-image sequences, on the left or right, was “more familiar”. On each practice trial, instead of the images seen previously, sequences of circles containing words or nothing appeared on either the left or right, one image at a time. The timing was matched with the exposure phase. A sequence of words indicated which side was the correct choice for that practice trial (e.g. “Choose”, “the”, “left”). Participants could not proceed past this practice phase until they consecutively and correctly responded to both practice trials. On each of 32 actual 2AFC trials, participants chose which of two 3-image sequences “felt more familiar” (Fig. 4b, third column). One sequence was a valid triplet, shown during exposure, and the other was a positional foil, created by combining first, second, and third position images from three different triplets (e.g. real triplet: C1, C2, C3; foil triplet: A1, B2, C3). Four foils were created in this way. During the test, the set of four real triplets and four foil triplets was fully crossed, for a total of 16 test pairs. An additional 16 test pairs were created by switching the order of presentation, such that each of the first 16 comparisons (left sequence vs right sequence) would 43 be shown right vs left instead, thus eliminating any left-right bias during test. The 32 trials were presented in a pseudo-random order. Participants initiated each trial by pressing the enter key. The first triplet was displayed on the left side of the screen, then there was an inter-triplet delay of 1000ms, and then the second triplet was displayed on the right side of the screen. Triplet type (“real” or “foil”) was counterbalanced such that each type occurred first and second an equal number of times. The dependent variable in this forced choice test was the percentage of correct triplets identified. Triplet Creation Task. In the third, and final, measure of learning, participants were informed that the 12 images seen earlier in the experiment had been composed of 4 triplets, with the images within each triplet always following one another in the same order. They were then asked to “create” the triplets, by selecting images they felt constituted a triplet. The 12 unique images were shown in a randomized 2 row x 6 column display. Beneath the image display, 12 drop-down boxes were spatially arranged into four groups of three. Each group of three dropdown boxes served as a placeholder for items in a triplet. Thus, by selecting which image should occur in each dropdown box, participants created and endorsed which images they believed constituted each triplet (Fig. 4b, fourth column). Triplet re-creation required both image identity and ordering to be correct. Each image could only be used once across all drop-down boxes, to ensure that all 12 images were assigned to a unique position. All 12 drop-down boxes required an entry before the participant could submit their selections and proceed to the post-experiment questionnaire. The dependent variable was the number of valid triplets re-created. 44 Post-Experiment Questionnaire. Following the three measures of learning, participants were asked a series of questions about their meta-cognitive state, use of strategies, and understanding of the experiment. Participants were asked to: assess their own performance on each of the three measures, assess if they knew there were four triplets prior to being told about the triplets in the instructions for the Creation task, indicate if they used any strategies in each of the phases (and if so, what strategy), and indicate if they had mentally assigned labels to any of the images and, if so, to list any labels they had assigned. Analysis 95% confidence intervals were estimated using a bootstrap procedure which, in short, generated data from 15000 theoretical experiments by sampling the observed data with replacement, then used the critical values at 2.5% & 97.5% from the resulting, sorted data as the confidence interval (see “Bootstrap Procedure” in Supplementary Materials). Descriptive Analyses Target Detection Task Analyses. The main dependent variable was “response time facilitation” (RTF), the response time difference between images in the first position of a triplet (i.e. less predictable images) and those in the third position of a triplet (i.e. perfectly predictable images). The observation of a facilitation between triplet positions, especially between the first 45 and third position (RTF1-3), is a standard indirect measure of learning and has shown that later images within a triplet are detected more rapidly than earlier images in a triplet (Turk-Browne et al., 2005; Kim et al., 2009; Musz et al., 2015). This variable was calculated both within and across participants. Response time differences between triplet positions. Comparisons of response times between triplet positions employed the Kruskal-Wallis H test as implemented in SciPy 1.3.2 (Jones et al., 2001). Since multiple tests were run (one for each condition), we corrected for multiple comparisons by applying a Bonferroni correction as implemented in statsmodels 0.10.2 (Seabold & Perktold, 2010). 95% confidence intervals were estimated using a bootstrap procedure (see “Bootstrap Procedure” in Supplemental Methods). Response time facilitation differences between conditions. Comparisons of response time facilitation between conditions employed the Kruskal-Wallis H test as implemented in SciPy 1.3.2 (Jones et al., 2001). Trend analysis on the facilitation data across conditions was performed using the Mann-Kendall test as implemented in PyMannKendall 1.4.1 (Hussain & Mahmud, 2019). 95% confidence intervals were estimated using a bootstrap procedure which generated multiple “surrogate” condition-specific datasets by sampling the observed data with replacement, then used the critical values at 2.5% & 97.5% of the resulting, sorted iterations as the interval (for details see “Bootstrap Procedure” in the Supplemental Methods). 46 Response Time Effect Size Estimate Analyses. For Kruskal-Wallis H tests, we converted the H statistic (as a proxy for 𝜒2) into an approximation of η2. The η2 approximation is shown in Eq. 1 (K. R. Murphy et al., 2014; Tomczak & Tomczak, 2014). The result of Eq. 1 multiplied by 100 “indicates the percentage of variance in the dependent variable explained by the independent variable” (Tomczak & Tomczak, 2014). Since this is an estimate, the calculation can sometimes produce a negative number (unlike η2), therefore, we report a value of “0” whenever the estimate is negative. (Eq. 1) 𝜂2 = 𝐻 − 𝑛𝑔𝑟𝑜𝑢𝑝𝑠 + 1 / (𝐻 + 𝑛𝑜𝑏𝑠𝑒𝑟𝑣𝑎𝑡𝑖𝑜𝑛𝑠 − 𝑛𝑔𝑟𝑜𝑢𝑝𝑠) Direct Measures of Learning (2AFC & Creation). Performance on the 2AFC task was assessed by calculating the mean percentage of correct choices made when assessing real triplet sequences vs. foil triplet sequences. Performance in the Creation task was assessed by calculating the mean number of triplets created. The distributions of 2AFC and Creation performance under the null hypothesis were estimated using Monte Carlo simulations (see “Monte Carlo Simulation” in the Supplemental Methods). Since our pre-registered hypotheses anticipated learning, we employed one-way tests to determine critical thresholds in the null distribution. P-values were computed by comparing observed means with critical thresholds in the null distribution. If performance in either task exceeded any of the thresholds (p< 0.05, 0.01 or 0.001), we reported the most unlikely threshold exceeded. 47 2AFC Analyses. Performance on the 2AFC task was assessed by calculating the mean percentage of correct choices made between previously shown, ordered groups of images (i.e. triplets) and group of images that had never been shown in that particular order (i.e. foils). Creation Analyses. Performance in the Creation Task was assessed by calculating the mean number of triplets created. In the Creation Task, a number from 1-12 was assigned to each image, randomly shown, in a display. Below the display of images were 12 drop-down boxes, organized into sets of three. The options in each drop-down box were image labels (1-12, as assigned in the image section). Participants were asked to create four triplets by selecting using the drop-down boxes. Relationships between measures. Spearman rank-order correlations (rs) were calculated using SciPy 1.3.2 (Jones et al., 2001). For Mann-Whitney U tests and Wilcoxon signed-rank tests, we used the Python package Pingouin 0.3.9 (Vallet, 2018) to calculate the rank-biserial correlation (RBC) or the matched pairs rank-biserial correlation (MPRBC; Kerby, 2014), respectively. Results Participants attended during the initial exposure to triplet structure 48 Participants in the 60 and 96 repetition conditions performed well in the “Jiggle” cover task during their initial exposure to the visual sequences. Stimulus movements (i.e. “jiggles”) were detected with 90% accuracy in both conditions, and the mean time to detect jiggles was less than 460ms in both conditions (Supp. Table 3). Participants in the “0 Repetitions” condition detected an equivalent number of targets during the Target Detection task (between 89-90%; Supp. Table 3), leading us to conclude that participants were attending to the stimuli during initial exposure and the TD task. Additionally, participants in the 60 and 96 repetitions conditions were required to confirm that they could “see each image” during the initial exposure to have their data analyzed. No evidence of learning in Target Detection Task We observed no evidence of learning via response time facilitation using our indirect learning measure (Figs. 5a-b). While previous studies have shown facilitation effects in response times (Turk-Browne et al., 2005; Kim et al., 2009; Bertels et al., 2012; Bays et al., 2015; see Experiment 3), we saw no differences between any of the triplet positions in any of the conditions. We observed mean RTF1-3 facilitation effects near zero, they were: 0.23ms (95% CI: [-7.48, 7.93]), 3.19ms (95% CI: [-4.42, 10.8]) and -1.09ms (95% CI: [-9.44, 7.26]) for the 0, 60, and 96 repetitions conditions respectively. A Kruskal-Wallis H test revealed no significant differences between response times by triplet position in any condition (H(2)<1.4 for all conditions; Table 2, “RT differences between triplet positions”). In addition to observing no differences in response times between triplet positions, we observed no monotonic decreases in response times by triplet position (i.e. response times did not quantitatively speed by subsequent triplet position). Since all 49 effect sizes were smaller than 4ms, and since the condition with most exposures (i.e. 96) provided no evidence of learning, we concluded that there was no meaningful learning-related speeding of response times. Overall, items that began a triplet were detected at the same speed as items that ended a triplet, regardless of predictability or the amount of exposure provided to the statistical regularities. Mixed evidence of learning using forced choice measure Forced choice accuracy was lower than previous reports (Fiser & Aslin, 2002; Schlichting et al., 2017), but numerically above chance in all conditions. Without correcting for multiple comparisons, performance exceeded statistical thresholds in two of the three exposure conditions (Fig. 5c). Forced choice performance was above chance in the “0 Repetitions” condition (mean=51.4%, 95% CI=[49.3, 53.4], p=0.045) and in the “96 Repetitions” condition (mean=51.7%, 95% CI=[49.6, 53.7], p=0.021 ), but not in the “60 Repetitions” condition (mean=50.2%, 95% CI=[48.1, 52.2]). However, neither of the uncorrected findings survived a Bonferroni correction for multiple comparisons (all corrected p-values > 0.05). Participants were able to re-create triplets Participants were able to create triplets above chance levels (critical threshold of mean triplets created=0.067) in all conditions (0 Repetitions: mean=0.15, p<0.001, CI=[0.060, 0.24]; 60 Repetitions: mean=0.13, p<0.001, CI=[0.067, 0.20]; 96 Repetitions: mean=0.15, p<0.001, 50 CI=[0.060, 0.24]; Fig. 5d-e; Table 2, “Mean triplets created”) and these findings held after applying a Bonferroni correction (all corrected p-values < 0.001). Participants also generated “foil” triplets (from the 2AFC task) at levels greater than chance (Table 1, “Mean foils created”; Supp. Fig. 2), suggesting that some participants learned the structure during the 2AFC test. However, performance on the Creation task appears to reflect true learning as: (a) real triplets were created more than foil triplets in each condition, and (b) a Wilcoxon signed-rank test of data combined from all conditions showed that triplets were created at a rate significantly higher than foils (W=572, p=0.030, MPRBC=-0.062). Direct measures were positively correlated, whereas indirect and direct measures had no relationship Consistent with our original hypothesis, the direct learning measurements (from the 2AFC & Creation tasks) were positively correlated with one another (rs = 0.15, 0.19, 0.20, for the 0, 60 and 96 repetition conditions; Fig. 5f; Table 2). The only non-significant correlation was in the “0 Repetitions” condition, where there was no initial exposure. Also as hypothesized, the direct learning measurements were uncorrelated with the indirect learning measurement, RTF1-3, in all conditions (Table 2; Supp. Fig. 3). 51 Table 2: Results of varying exposure (Experiment 1) Attend Jiggle (0 Repetitions) Attend Jiggle (60 Repetitions) Attend Jiggle (96 Repetitions) RT differences between triplet positions H(2) = 1.01, η2=0 H(2) = 1.37, η2=0 H(2) = 1.12, η2=0 Mean RTs by triplet position (ms) (1st) 424.6 (2nd) 424.9 (3rd) 424.2 (1st) 412.2 (2nd) 412.8 (3rd) 409.1 (1st) 414.8 (2nd) 416.1 (3rd) 416.5 Mean RTF1-3 (ms) + 95% CI 0.23; [-7.48, 7.98] 3.19; [-4.42, 10.8] -1.09; [-9.44, 7.26] Mean % correct (2AFC) + 95% CI 51.4 [49.3, 53.4] 50.2; [48.1, 52.2] 51.7 [49.6, 53.7] Mean triplets created 0.15, p<0.001; [0.060, 0.24] 0.13, p<0.001; [0.067, 0.20] 0.15, p<0.001; [0.060, 0.24] Mean foils created 0.083, p<0.001; [0.076, 0.090] 0.083, p<0.001; [0.078, 0.088] 0.075, p<0.001; [0.070, 0.080] Difference between foils and real triplets created W=46.5, MPRBC= -0.069 W=87.5, MPRBC=-0.081 W=66.0, MPRBC=-0.055 Spearman rho (rs, 2AFC & Creation) 0.15 0.19, p=0.035 0.20, p=0.026 Pearson r (RTF1-3 & 2AFC) 0.046 -0.046 -0.12 Spearman rho (rs, RTF1-3 & Creation) 0.053 -0.037 -0.0060 Notes. H statistics are calculated with Kruskal-Wallis H tests between groups (see “Effect Size Analyses”). Eta squared estimate (η2) is calculated following Equation 1. W statistics are calculated with Wilcoxon signed-rank tests (see “Effect Size Analyses”). Matched pairs rank biserial correlation (MPRBC) is a measure of effect size (see “Effect Size Analyses”). 52 Figure 5: Results from manipulating amount of repetition during the Jiggle Task. (a) Mean RTs by triplet position and condition. (b) Response time facilitation effects between triplet first and third positions. (c) Percentage of “real” triplets successfully chosen during 2AFC task. Gray dots indicate a single participant. Red dots indicate mean of distribution. (d) Mean number of triplets created (out of 4 possible). (e) Histogram showing the number of participants creating triplets, from none (0) to all (4). (f) Spearman correlation between direct measures (2AFC & Creation). Colored dots represent individual participants. Black dots indicate means (when more than four data points are available). (a-d, f) Vertical line segments indicate 95% CIs. (b-d) Dashed gray lines indicate chance performance. Repetition did not affect learning Across all measures of learning we observed no significant difference in performance between the exposure conditions (Figs. 5a-e). We observed no differences between conditions on any learning metric (RTF1-3: H(2)=0.88, η2=0; 2AFC: H(2)=1.52, η2=0; Creation: H(2)=0.039, η2=0). Thus, exposure amount did not affect response time facilitation (RTF1-3), forced choice performance (2AFC), or triplet creation (Creation). We had predicted that increasing the amount of initial 53 exposure to regularities would benefit performance in the direct tasks (2AFC & Creation) while having little effect on the indirect measure (RTF1-3). However, we did not observe monotonically increasing performance in any of the tasks as a function of the amount of initial exposure to regularities. Discussion The accuracy of our pre-registered hypotheses was mixed. Conflicting with our pre-registered hypotheses, we found no evidence of visual statistical learning in the target detection (TD) task and we observed very weak evidence of learning in the 2AFC and Creation tasks. Unlike prior work (Fiser & Aslin, 2002; Turk-Browne et al., 2005), we found no evidence that increased exposure to statistical regularities significantly improved performance on any measure of statistical learning. Notably, participants with zero exposure to the triplets prior to test demonstrated equivalent performance to participants who had more than 15 minutes of exposure in the Jiggle task (Fig. 5). In accordance with our pre-registered hypotheses, we observed a correlation between the direct measures (2AFC and Creation). The lack of correlation between the direct measures and the indirect measure (TD) was consistent with our hypothesis that statistical learning reflects a mixture of distinct processes. However, because no learning was observed using the response time facilitation metric, it remains unclear how strongly this evidence could be taken to support the theory of distinct learning processes. 54 Our observation of overall weak learning, with no effect of manipulating the amount of exposure, could indicate that performance was hampered by our constellation of experimental parameters (e.g. stimulus type, method of initial exposure, specific instructions). We therefore investigated the method of initial exposure in a follow-up experiment (Exp. 2), hypothesizing that learning would improve across all measures. Experiment 2: Cover task modulated visual statistical learning The genesis of the subfield of “statistical learning” began with the striking observation that infants can segment repeating sequences of nonsense syllables after only a few minutes of exposure (Saffran, Aslin, et al., 1996; Frost et al., 2019). Both children and adults can learn the associations between consecutive sounds, even if the sounds are played in the background, while participants are doodling (Saffran et al., 1997). These early findings led to the belief that statistical learning was automatic. The original statistical learning paradigm has since been extended in two significant ways. First, researchers began studying learning in other modalities, especially in vision (Kirkham et al., 2002; Fiser & Aslin, 2002). Second, the statistical regularities began to be presented using a diverse set of cover tasks. The methods for presenting visual regularities range from passive exposure (Kirkham et al., 2002; Fiser & Aslin, 2002; Kim et al., 2009) to 1-back tasks (Turk-Browne et al., 2005; Brady & Oliva, 2008; Arciuli & Simpson, 2012; Campbell et al., 2012), and to tasks involving motion detection (Turk-Browne et al., 2009), image categorization (Turk-Browne et al., 2010; Vickery et al., 2018), image orientation (Zhao et al., 2013), background change detection (Arciuli 55 et al., 2014), image detection (Bays et al., 2015), or even direct instruction about the regularities (Siegelman, Bogaerts, Kronenfeld, et al., 2018). It is scientifically important that statistical learning is automatic, because this allows the same learning mechanism to be studied across the diverse task settings noted above. Moreover, automaticity of learning is one of the essential features of the psychological construct. In the framing of Saffran et al. (1999), “this learning process proceeds automatically as a byproduct of mere exposure.” Important early studies of automaticity observed that selective attention to the training stimuli was necessary for statistical learning (Toro et al., 2005; Turk-Browne et al., 2005). The claim was that, provided the stimuli are selectively attended, the statistical learning process operates automatically. Such an automatic learning process seems adaptive, because learning the recurring associations between co-occurring events, places, and other stimuli in our environment should facilitate faster and more accurate behavior. In Experiment 1, we found that regardless of the number of repetitions (up to 96), attention to stimulus motion (Turk-Browne et al., 2009) was insufficient to produce learning. These findings suggest that visual statistical learning is not automatic. We therefore decided to investigate why our findings differed from the broader field of visual statistical learning, focusing on the choice of cover task. Since the Jiggle Task was meant to discourage participants from explicitly searching for regularities in a stream of images, perhaps it limited their ability to learn regularities. We hypothesized that in contrast to the belief that passive exposure to regularities would be absorbed via “mere exposure” (Aslin, 2017), learning would improve as a result of goal state and task requirements of processing the stimuli (Vickery et al., 2018). 56 Another possible explanation for why we observed such poor performance in Experiment 1 is that our procedures for measuring learning were flawed in some way. Therefore, to provide a positive control for learning measurement, we designed a reference learning condition (“Feedback Training”). In the reference condition, we replaced the “Jiggle” cover task with a shorter exposure phase which explicitly informed participants about the triplet structure. Additionally, we provided feedback to the participant on their accuracy in a triplet creation task, until the participant demonstrated explicit knowledge by re-creating all four triplets simultaneously. We reasoned that explicit knowledge should improve performance (at least on the direct measures) and establish a positive control for our measures. This procedure not only allowed us to validate our measures, but also allowed us to determine which of our dependent measures (TD, 2AFC, and Creation tasks) can reflect learning when participants have explicit knowledge of the regularities. In Experiment 2, we manipulated goal and attentional states by assessing visual statistical learning across five different statistical learning conditions and one reference condition. Of these six conditions, the reference condition (“Feedback Training”) was considered an “intentional” learning task because participants were given the explicit goal to learn the triplets. Similarly, two of the statistical learning conditions were considered intentional, as they gave participants a goal related to statistical structure (“Attend Relations”, “Attend Triplets”). The final three conditions were hypothesized to rely upon incidental learning (“Attend Stimuli”, “1-back”, “2-back”) (see “Cover Tasks” below). We observed that participants’ overt task goals strongly modulated learning. Overall, our data suggest that standard statistical learning paradigms have not been measuring an automatic 57 process that is simply gated by selective attention; they measure a combination of learning processes, heavily dependent on how participants process the relationships between stimuli. Method Participants. Participants (n=654, ages 18-74) performed visual sequence learning in one of six conditions (five statistical learning conditions, and one “Feedback Training” condition). Demographic information is provided in Suppl. Fig. 1. We analyzed data from 120 participants in each of the statistical learning conditions, and 54 participants in the “Feedback Training” condition. Data were collected online using Amazon’s Mechanical Turk platform, coordinated using PsiTurk (T. M. Gureckis et al., 2016). Participant compensation varied between $4.40 and $5.75, depending on the approximate duration of each condition (Suppl. Table 1). To ensure high data quality, we limited participation to “US only” participants who had completed at least 99% of their prior Amazon tasks successfully. Exclusion Criteria. Because the cover task was a key independent variable in these studies, and because our claims depend on ensuring task engagement, it was critical to ensure high levels of task performance. Therefore, we applied a set of 10 pre-registered exclusion criteria (https://osf.io/v56zx; Suppl. Table 1; Kraut et al., 2004). In addition, participants were prevented from participating if their Amazon Mechanical Turk ID had participated in a previous condition or pilot of this experiment. 58 Under these strict performance criteria, exclusion rates ranged from 29% to 77% (Suppl. Table 1). Especially high exclusion rates were seen in the 2-back condition, for which data was collected in the early months of the COVID-19 pandemic (as was the 1-back condition), but the data is consistent with both our hypotheses and other conditions collected at other time points; our conclusions do not depend on these data. Stimuli. Identical to Experiment 1 (Fig. 4a). Procedure The procedure was identical to that of Experiment 1 (Fig. 4b), except where noted. All conditions of the experiment consisted of an exposure phase (in which participants were exposed to visual sequential regularities) followed by a test phase (in which we indirectly and directly assessed their knowledge of the regularities). Exposure Phase. During exposure, participants viewed a stream of images. The stream was composed of pseudo-randomly ordered “triplets” (3 images in a fixed order). Each image stream employed 4 triplets, composed of 12 distinct images (Fig. 4a; Schapiro et al., 2012). The same 12 images were used to generate the image streams across all participants, but the order of images differed across participants. In the statistical learning conditions, image streams were composed of 720 images (i.e. 60 presentations of each triplet); in the Feedback Training condition, image streams were between 8-56 images (depending upon how quickly the participant learned; see “Feedback Training” below). Images in all but one condition were shown for 800ms with a 200ms 59 inter-stimulus interval (ISI). In the “2-back” condition, we lengthened the ISI to 1200ms, as our pilot data suggested participants struggled with the cover task at the faster rate (Suppl. Table 2). The assignment of images to triplets (i.e. the “stimulus set”) differed for each participant in all the statistical learning conditions. The same collection of 120 “stimulus sets” from Experiment 1 was used across all statistical learning conditions. Each stimulus set was counterbalanced to ensure that images appeared equally often in each triplet position (i.e. equally often in the first, second, and third position). We manipulated participants’ attentional and goal states during image sequence exposure by assigning them to perform a cover task. Assignments to cover tasks was a between-subjects manipulation, and each participant performed exactly one cover task. Cover Tasks. Below is a brief description of each cover task condition and any significant differences between them. More detailed instructions and experimental code for each task can be found on OSF (https://osf.io/yrq8t/). Note that the key manipulation in Experiment 2 was replacing the “Jiggle” cover task from Experiment 1 (Fig. 4b, leftmost column) with the conditions listed below. Statistical learning conditions: 1) 1-back (Supp. Fig. 2): Participants were instructed to watch a stream of images and to detect any images that were repeated back-to-back (e.g. X – R – R – Y; Turk-Browne et al., 2005). Participants were instructed to press the space bar within 800ms of a repeat 60 occurring. Participants practiced the 1-back task with black geometric shapes in two stages: a first stage with feedback on correct detection, then a second stage without feedback. The second stage was repeated until all three 1-back repeats were caught. 2) Attend Stimuli: Participants were instructed to watch a stream of images and to pay attention to each image, as they would be asked questions about them later (Fiser & Aslin, 2002). Participants performed no practice phase since their task did not require any overt action. 3) Attend Relations: Participants were instructed to watch a stream of images and to “pay attention to each image and when it appears in relation to other images”. They were also told to “learn which images are grouped together, [focusing] on which images appear before or after one another”. Participants performed no practice phase since their task did not require any overt action. 4) 2-back: Participants were instructed to watch a stream of images and to detect any images that were repeated after one intervening image (e.g. R – X – R; Vickery et al., 2012). Participants were instructed to press the space bar within 2000ms of a 2-back repeat occurring. Participants practiced the 2-back task with black geometric shapes in two stages: a first stage with feedback, then a second stage without feedback. The second stage was repeated until all three 2-back repeats were caught. 5) Attend Triplets: Participants were instructed to watch a stream of images where the images occur in groups of three. They were instructed that there would be four such groups, and their goal was to learn the order of the images in each of the four groups. Participants performed no practice phase since their task did not require any overt action. 61 Reference learning condition with explicit instructions and feedback: 6) Feedback Training: Participants were instructed to watch a stream of images, where the images occur in groups of three (i.e. triplets) and that there were four such triplets. They were told their goal was to “quickly and accurately” learn the four sets of triplets as well as the ordering within each triplet. Participants were shown each of the four triplets, with the same timing as the statistical learning conditions, followed by a 1000ms pause after the fourth triplet was shown. Then each of the four triplets were shown a second time. Participants were suggested to name the images or come up with other mnemonics to better learn the groupings. After seeing each triplet twice, participants attempted to re-create all four triplets (see Exp. 1, “Creation Task”). If they successfully created all four triplets, they moved on to the test phase. Otherwise, they saw the same exposure phase again (i.e. each triplet shown twice, following by the Creation Task). Participants were limited to 10 attempts, but all participants successfully re-created the triplets before the threshold. Feedback Training Differences. The exposure phase and test phase for the feedback training condition were identical to the statistical learning conditions, except for the following differences. First, only 54 participants were assigned to this condition. Second, the exposure phase was limited to between 2-20 presentations of each triplet. Third, only the first 18 stimulus sets (assignments of images to triplets) from the statistical learning conditions were used, with each stimulus set seen by 3 participants. Finally, as described above, the cover task involved direct 62 feedback. Test Phase (Measures of Learning). After the exposure phase, participants completed three ordered tests of learning (Fig. 4b, columns 2-4): (1) target detection, measuring image response times; (2) a two-alternative forced choice (2AFC) between triplets; and (3) explicit re-creation of triplets in a matter identical to Experiment 1. Analysis Pre-registration: After data collection and analysis of Experiment 1, we pre-registered additional predictions for conditions: “Attend Stimuli”, “Attend Relationships”, and “Attend Triplets” using the AsPredicted template (https://osf.io/v56zx). Power Analysis: The rationale for our sampling procedure was almost identical to Experiment 1. The only condition using a smaller sample size was the “Feedback Training” condition (n=54), which served as a comparison for ceiling performance. Descriptive Analyses: Identical to Experiment 1. Results 63 Participants recognized and recreated visual regularities We observed evidence of learning in all six conditions using the 2AFC task (Figure 6a) and the Triplet Creation task (Figure 6b). These learning effects remained statistically significant after correcting for multiple comparisons (all significant p-values <0.001; Table 3). Participants in the “Feedback Training” condition were, on average, able to re-create all four triplets after 3 attempts during the exposure phase (median=3, mean=3.07; Supp. Table 3), with no participant requiring more than 7 attempts to achieve perfect creation accuracy. Attentional and goal states robustly affected learning Forced-choice familiarity performance was significantly different across learning conditions (2AFC: Kruskal-Wallis H=75.0, p<0.001, η2=0.097; Fig. 6a). Participants’ ability to re-create the triplets following exposure was also significantly different across the learning conditions (Creation: Kruskal-Wallis H=180, p<0.001, η2=0.21; Fig. 6b). Given the significant differences between conditions, we ran post-hoc Mann-Kendall trend analyses to test whether cover tasks involving more attention to temporal regularities (as ordered in Figure 6, left to right) were associated with increased learning. In support of our hypothesis, we observed increasing trends in both tasks (same trend in 2AFC & Creation: τ=1.0, z=2.63, p=0.0085). 64 Participants who re-created all the triplets performed well on the 2AFC task, while participants who created zero triplets performed near chance levels on the 2AFC task We found that when participants could create all 4 triplets, their average 2AFC performance was near ceiling (Table 3, “2AFC % correct by x triplets created – 4”; Fig. 6c). This might be expected, as participants who have knowledge of the triplets should be able to exploit that knowledge in the 2AFC task as well. However, statistical learning has long been described as an “implicit” learning process (Perruchet & Pacton, 2006). Therefore, we calculated the 2AFC performance amongst those participants who were unable to explicitly re-create any triplets. We found that, for participants who created 0 triplets correctly, their average 2AFC performance was at or near chance (Table 3, “2AFC % correct by x triplets created – 0”; Fig. 6c). Additionally, we observed a significant and positive correlation between the two direct measures (Forced Choice and Triplet Creation) in all conditions except “Feedback Training” (Fig. 6c). Collectively, these data are compatible with the idea that the performance in the Forced Choice task and the Triplet Creation task arise from the same underlying knowledge. Measures of response time provided no evidence of learning Response time facilitation effects were not consistently observed (Fig. 6d) and differed significantly between the six conditions (RTF1-3: Kruskal-Wallis H=15.8, p=0.0075, η2=0.016; Fig. 5e). Within conditions, three showed numerically negative facilitation and three showed numerically positive facilitation (Figure 6e; Table 3). We observed no response time facilitation 65 trends between conditions after conducting a post-hoc Mann-Kendall trend analysis (τ=-0.067, z=0.0). There was little evidence that more predictable images were detected faster than less predictable images. The two largest RTF1-3 effects were observed in the “Attend Relations” condition (uncorrected p = 0.039, corrected p = 0.23) and the “2-back” condition (uncorrected p = 0.028, corrected p = 0.17). Importantly, our largest response time facilitation of 12.4ms is less than half the size of the previously reported facilitation effects (e.g. 35ms (Bays et al., 2015), 38ms (Bertels et al., 2012), and 46ms (Kim et al., 2009)). As in Experiment 1, RT facilitation did not exhibit any correlation with either of the other measures of learning (Suppl. Fig. 3). 66 Table 3: Descriptive statistics for Experiment 2. All p-values are Bonferroni-corrected (except rs values). The “2AFC % correct by triplets created” means are calculated only with sufficient data (4+ participants). 1-back Attend Stimuli Attend Relations 2-back Attend Triplets Feedback Training n 120 120 120 120 120 54 Mean 2AFC (95% CI) 54.0 (51.7, 56.4), p<0.001 57.1 (54.1, 60.1), p<0.001 62.3 (58.6, 66.1), p<0.001 66.9 (63.3, 70.5), p<0.001 66.9 (63.2, 70.7), p<0.001 81.3 (76.2, 86.4), p<0.001 Mean triplets created (95% CI) 0.15 (0.046, 0.25), p<0.001 0.32 (0.19, 0.45), p<0.001 0.74 (0.50, 0.97), p<0.001 0.94 (0.69, 1.20), p<0.001 0.97 (0.72, 1.21), p<0.001 3.30 (2.97, 3.63), p<0.001 Spearman correlation (rs; 2AFC vs Creation) 0.26, p<0.01 0.44, p<0.001 0.44, p<0.001 0.74, p<0.001 0.77, p<0.001 0.02 # (%) participants creating x triplets 0 108 (90.0) 94 (78.3) 84 (70.0) 69 (57.5) 66 (55.0) 2 (3.7) 4 2 (1.7) 2 (1.7) 14 (11.7) 17 (14.2) 15 (12.5) 38 (70.4) 2AFC % correct by x triplets created 0 52.6 52.9 54.9 54.8 52.8 n/a 4 n/a n/a 94.1 94.5 92.1 80.6 Response times by triplet position (ms) 439.1, 441.6, 443.7 428.9, 435.7, 428.2 428.6, 422.3, 419.9 439.3, 431.0, 428.1 427.8, 421.4, 432.7 411.5, 422.9, 421.7 Mean response time facilitation 1-3 (ms) (95% CI) -5.09 (-13.3, 3.13) 1.13 (-8.46, 10.7) 9.85 (-0.36, 20.1) 12.4 (3.05, 21.8) -4.74 (-14.0, 4.54) -10.7 (-23.1, 1.82) 67 Figure 6: Experient 2 results. (a) Performance of each subject on the 2AFC measure (% of correct choices made), by condition. (b) Histograms of the mean number of triplets successfully re-created, by condition. (c) Relationship between performance on the 2AFC task and the Triplet Creation task. rs indicates Spearman’s rank-order correlation. (d) Mean RTs by triplet position and condition. (e) Response time facilitation effects between triplet first and third positions.. Vertical line segments indicate 95% confidence intervals (where there are at least 4 data points). Dashed gray lines indicates chance performance. 68 Discussion Confirming our pre-registered hypotheses, and relative to Experiment 1, we found that making small changes to the instructions or to task goals could dramatically increase visual statistical learning. These findings not only validate that our direct measures are valid assessments of learning, but also suggest that research in the field of statistical learning has not sufficiently grappled with the active engagement of the learner during the learning process. Most statistical learning research assumes a passive learning process; one in which associations are created between any and all repeated stimuli, as long as they are attended. This data suggests that there exists a gradient of learning, with cursory attention to stimuli being insufficient for robust learning (Exp. 1, Fig. 5c) and attention to temporal relationships between stimuli vastly improving learning (Exp. 2, Fig. 6a). Choice of cover task affected visual statistical learning Passive attention to stimuli and 1-back tasks are the standard methods for exposing participants to regularities in visual statistical learning experiments (Fiser & Aslin, 2002; Kirkham et al., 2002; Turk-Browne et al., 2005; Brady & Oliva, 2008; Arciuli & Simpson, 2012; Stevens et al., 2014; Musz et al., 2015; Siegelman, Bogaerts, Kronenfeld, et al., 2018). As participants focus more on temporal aspects between stimuli, we hypothesized that visual statistical learning would improve. This is what we observed. Using the 2AFC performance to compare between experiments and conditions, we found that when participants attended to perpetual characteristics unrelated to the stimuli (Exp. 1; Fig. 5c), learning was near chance levels. When the focus was on individual stimuli (“Attend 69 Stimuli”, “1-back”), performance improved slightly (Fig. 5a). As the focus shifted to more temporal aspects, performance improved even more (“Attend Relations”, “2-back”, “Attend Triplets”; Fig. 6a). Finally, our reference condition (“Feedback Training”) revealed that participants could explicitly learn the triplet structure, performing near ceiling (Fig. 6a). How does the learning performance compare to previous studies? While we observed lower 2AFC results than some studies (e.g. Fiser & Aslin, 2002; Schlichting et al., 2017), our results were in-line with other large-sample studies (Arciuli & Simpson, 2012; Vickery et al., 2018). As noted above, attentional demands and goal states seem to impact the degree to which statistical learning occurs, and so differences across studies could arise from small changes in participants’ task orientation or expectations during the exposure phase. This is consistent with previous work comparing categorization tasks with more perceptual tasks, where the choice of cover task modulated learning (Vickery et al., 2018). Though we observed no evidence of learning in response time facilitation, we confirmed with informal in-lab tests that participants could theoretically use explicit knowledge of the sequences to respond more quickly to targets, producing very large (>100ms) response time speeding for the target item. However, we observed no evidence of participants pursuing such a strategy. Thus, our response time facilitation measure was largely insensitive to explicit knowledge of the triplet regularities. The lack of learning in our Target Detection task is investigated more deeply in Experiment 3. 70 Robust learning was observed in the direct measures within the “Feedback Training” condition We observed robust learning in the direct measures in our reference condition (“Feedback Training”; Figs. 6a-b). This provided a positive control for our 2AFC and Creation testing paradigms. Thus, the very weak evidence of learning in the direct measures in Experiment 1 most likely reflect properties of the incidental exposure method (i.e. Jiggle Task) that was used. Although learning in the “Feedback Training” condition was strong (with 50% of participants achieving > 90% 2AFC), one-third of participants did not achieve 2AFC accuracy >=75%. Previous work has shown that around one-third of participants demonstrate chance performance in statistical learning tasks (Siegelman & Frost, 2015), so it is possible that a subset of individuals was unable to express learning under standard 2AFC testing. Chapter Discussion Visual statistical learning did not occur automatically Our results suggest that visual statistical learning (VSL), as measured using the standard paradigm deployed here, is not automatic. VSL did not consistently and robustly arise from mere exposure, even when attending to the stimuli. We found little evidence of learning when participants performed a task that required selective attention to the stimuli without requiring any temporal processing (Fig. 5). Conversely, we observed robust learning when the task required participants 71 to compare and relate stimuli across time (e.g. “2-back”; Fig. 6). Collectively, we found that the specific kinds of attentional demands and goal states employed by participants strongly modulated their statistical learning (Figs. 5-6). These results do not suggest a continuously operating, incidental learning process, independent of other cognitive processing (Hasher & Zacks, 1979), but rather intention and cognitive demands significantly influenced how well regularities were learned and whether they were learned at all. This is in contrast to the view that if participants are exposed to visual sequences containing simple repeated regularities, and if they selectively attend to the stimuli, then statistical learning will occur (Fiser & Aslin, 2002; Turk-Browne et al., 2005; Aslin, 2017). A continuously operating, “automatic” learning process can only benefit from increased exposure. Thus, the results of Experiment 1 contrast with the automaticity view because increasing the amount of exposure to regularities, from no (0) triplet repetitions, to a moderate number of repetitions (60), to a high number of repetitions (96), had no effect on learning. It could be that these results were specific to usage of the “Jiggle” cover task (i.e. when the cover task did not require temporal or stimulus-specific processing), but they nonetheless suggest that visual statistical learning is not as automatic as previously believed. Previous work in both vision and audition has suggested that selective attention to the stimuli in statistical learning paradigms is necessary for learning regularities (Turk-Browne et al., 2005; Toro et al., 2005). While there has been some disagreement about whether selective attention is strictly necessary (Musz et al., 2015), there is agreement that when selective attention is deployed, then statistical learning should occur. The implication of prior work is that selective attention is necessary and sufficient for statistical learning to occur. Specifically, that selective attention should 72 lead to automatic VSL (Turk-Browne et al., 2005). Our results suggest this is not the case since we observed little evidence of statistical learning despite attention to the stimuli (Exp. 1). In contrast, the higher performance in Exp. 2 relative to Exp. 1 suggests that learning was strongly modulated by cover task. Despite the similarity that participants attended to the stimuli in all conditions, almost all of the learning performance could be eliminated by changing the task that participants performed. Therefore, it appears that the learning processes operating in standard visual statistical learning paradigms are largely not automatic, but rather modulated by the goal-state and attentional state of the participants. As this traditionally studied learning process is more volitional and attention-dependent than had been supposed, future work should investigate whether these results are specific to the visual domain. More generally, learning researchers should aim to better distinguish the kinds of learning that occur in statistical learning paradigms and in other, more conventional, paradigms, such as paired associate learning. More robust performance across participants was expected with an automatic learning process An automatic, “always on” learning process was expected to transpire for all participants. Since much of the learning from the direct measures was only nominally above chance, it does not provide strong evidence for an automatic learning process. Even a small subpopulation of participants who (for whatever reason) noticed any structure, could have driven almost all of the group learning effects. For example, in the “60 Repetitions” condition (Exp. 1), there was a single 73 participant (of 120) who recreated all of the triplets. Removing this one participant from the condition reduces the mean 2AFC performance below 50%. Moreover, most participants were unable to recreate any triplets in the Creation task (Table 3, “# (%) participants creating x triplets -- 0”). The 2AFC performance amongst such participants (who could not re-create a triplet) was less than 55% in all conditions (Table 3, “2AFC % correct by x triplets created – 0”). Thus, a minority of participants with explicit knowledge (possibly arising from idiosyncratic approaches to the task) can account for almost all of the learning in these paradigms (Fig. 10; see also de Leeuw, 2016). The association between explicit knowledge and learning is investigated further in Chapter 3 (“Explicit knowledge was evident in participants who exhibited the best learning”). Can learning be automatic? Our findings do not preclude the existence of automatic learning, or that there may be automatic forms of statistical learning that operate over longer timescales than those measured in the lab (Henke, 2010; P. J. Reber, 2013). In fact, we believe that there certainly are some forms of automatic regularity learning (e.g. Chait, 2020) and participants may perform some of the learning implicitly. While it is unclear how much “mere exposure” contributes to rapid learning of visual sequential relationships, it is evident that examples of presumably unintentional learning do exist, especially in the auditory modality (Aslin, 2017). However, in our visual statistical learning paradigm, learning does not seem to be automatic. Method of stimulus engagement strongly modulated learning 74 The automaticity view assumes a passive learning process in which associations are created between any and all repeated stimuli, as long as they are attended. As such, statistical learning research tends to use a standard paradigm of familiarization to regularities followed by measure(s) of learning. However, the method by which researchers expose participants to regularities can vary widely. The diversity of statistical learning cover tasks was assumed to have little effect upon the resultant learning, because the learning was occurring automatically and, presumably, was not much affected by cover task. Our data (Figs. 5c-e; 6a-b) suggest that statistical learning research has not sufficiently grappled with the active engagement of the learner (i.e. choice of cover task), and with the knowledge participants bring to bear during learning. Scientific conclusions are often drawn under the assumption of almost absolute naiveté from participants, regardless of the degree of attention to temporal regularities, contextual cues in the instructions, or the fact that the participant knows they are in a Psychology experiment (e.g. Saffran, Newport, & Aslin, 1996; Siegelman, Bogaerts, Elazar, Arciuli, & Frost, 2018). Subtle differences in instructions or cover tasks can dramatically alter the learning process and the resultant knowledge. Our findings suggest that tasks requiring an emphasis on relational processing (e.g. Turk-Browne et al., 2005; Brady & Oliva, 2008) will improve visual statistical learning. We believe that a critical aspect of visual statistical learning is the method of engagement with the regularities which, in turn, partially determines the extent to which statistical learning occurs. 75 Learning assessed by 2AFC and Creation tasks may share a learning process and/or memory store The relationship between performance on the direct measures (2AFC and Creation) was positively correlated in all conditions, while the indirect and direct measures remained uncorrelated in all conditions. Consistent relationships between the direct measures suggest a shared underlying learning process and/or memory store. While it is unclear whether the 2AFC task assesses knowledge “implicitly” (based upon recognition as opposed to recall), the Creation task can only be performed by actively selecting and ordering elements from a set of 12 images and is therefore, more “explicit”. This implicit/explicit distinction may also be irrelevant as the memory store may be identical for both processes (Reder et al., 2009). This topic is explored more deeply in Chapter 3. Chapter 2 Summary The findings in Chapter 2 support the claim that when participants were not focused on temporal relationships between stimuli, then those relationships were weakly learned or not learned at all (Exp. 1). In contrast, participants robustly learned the regularities when attending to temporal aspects of the stimulus relationships (Exp. 2). 76 Chapter 3: Limited evidence of implicit visual statistical learning Portions of this chapter were posted on PsyArXiv (Himberger et al., 2019). How implicit is statistical learning? The definition of statistical learning typically specifies that it occurs without intent and produces implicit knowledge. Finding that statistical learning occurs with explicit intent does not contribute to the theory that statistical learning is a distinct form of learning. Similarly, finding that the resultant knowledge is explicit would suggest that “explicit statistical learning” or “intentional statistical learning” may better be classified under standard associative learning (Mayes et al., 2007). Therefore, in this chapter, we focus on one of the key assumptions of statistical learning research, which is that regularities are implicitly learned and produce implicit knowledge. As shown in Chapter 1, “implicit learning” research pre-dates the field of “statistical learning” (A. S. Reber, 1967; Schacter, 1987; Saffran, Aslin, et al., 1996). However, statistical learning and implicit learning are often thought to rely upon the same underlying learning processes (Pierre Perruchet & Pacton, 2006; Stevens et al., 2014; c.f. Batterink et al., 2019). Others suggest a more complicated picture, with statistical learning occurring both implicitly or explicitly and generating a combination of implicit and explicit knowledge (Batterink, Reber, Neville, et al., 2015; Batterink, Reber, & Paller, 2015). Adding to the confusion, one group found no benefit of explicit instruction relative to incidental learning (Arciuli et al., 2014), while the same group found no learning with concurrent motor requirements unless instructed to learn the regularities (Stevens et al., 2014). 77 This variety of opinions may result from the imprecise use of the word “implicit”. Implicitness can refer to the learning process, the resultant knowledge (sometimes called “implicit memory”), or even the instructions (R. Ellis, 2009). Ellis (2009) suggests that any one or all three of these aspects (learning, knowledge, and instruction) can be either explicit or implicit (Table 4). Table 4: Aspects of implicit & explicit: instruction, learning, and knowledge (content from Ellis, 2009). Instruction Learning Knowledge Implicit Instructions which do not concern to-be-learned content Learning without intent and conscious awareness Non-declarative access to learned content Explicit Instructions about content to learn Learning with intent or conscious awareness Declarative access to learned content Concretely, implicit or explicit instruction can lead to implicit or explicit learning. Both implicit and explicit learning can lead to implicit or explicit knowledge. Implicit knowledge is defined by the inability to verbalize said knowledge, but with the ability to otherwise demonstrate that the learning occurred. Explicit knowledge is the most recognizable and includes facts and episodes from our lives in the form of “declarative memory” (Fig. 1; Squire & Zola, 1996). While statistical learning is characterized as “implicit”, it does not always fall neatly under the implicit distinctions in Table 4. For example, in Experiment 1, participants were not instructed about the regularities and hypothesized to implicitly learn through “mere exposure” (Aslin, 2017). After exposure, they exhibited no evidence of learning using an indirect metric (response time facilitation) meant to assess implicit knowledge (Figs. 5a-b). However, participants who recognized visual sequence regularities in a forced-choice task could also often recreate the sequences when explicitly probed, suggesting an explicit, conscious form of knowledge (Fig. 5f). 78 Another possible source of confusion regarding the implicitness of statistical learning is in the choice of cover task. In Chapter 2, we measured visual statistical learning (VSL) under six exposure conditions and characterized the relationship between three different measures of learning. Across all conditions, we found no evidence of VSL from the indirect measure, but some evidence of VSL in the direct measures; Fig. 6). This is particularly striking because the indirect measure is thought to assess implicit knowledge, whereas the direct measures are thought to assess more explicit knowledge. Beyond the differences in resultant memory, a crucial difference between Experiments 1 & 2 was that some of the cover tasks used during exposure to the regularities were biased towards a more explicit form of learning. In the most extreme example, participants were told that four triplets existed and they would need to learn them quickly and accurately (“Feedback Training”). Participants then trained on creation of the four triplets until they were able to recreate all four in the same attempt. This task is clearly a form of explicit learning by the standard definition. These findings suggest that the instructions preceding the exposure phase can shift learning to become either implicit or explicit. Additionally, the choice of cover task itself can shift learning towards explicitness, even in an ostensibly incidental, implicit learning paradigm. Thus, as suggested by Ellis (2009), there is not a 1-1 relationship between implicit/explicit instructions and implicit/explicit learning. For example, the “Attend Triplets” condition involved intentional learning, which could be categorized as both explicit instruction and explicit learning. More ambiguously, the “Attend Relations” condition involved explicit instruction, but was less proscriptive, as participants were instructed to attend to the relationships between stimuli. The instructions for the three other SL conditions from Exp. 2 were less proscriptive still and this shifts 79 the instructions categorically from explicit to implicit. Similarly, the change in instructions could shift the learning process to implicit, though this is more difficult to assess. Despite the binary nature of these definitions (Table 4), there is likely a gradient between the categories and not a strict boundary. In this chapter, I investigate two ways in which visual statistical learning is not as implicit as previously conceptualized. First, in Experiment 3, we show that the typical indirect measure used to demonstrate implicit visual statistical learning suffers from a methodological issue in assessing response time facilitation. We demonstrate how response time facilitation can arise without any learning, specifically finding that response times decrease for targets that occur later within test trials, and the response times for later items within a trial are used as the index of learning. Second, we analyze responses from participants in Experiments 1-3 to assess if explicit knowledge is associated with the learning measured in the direct tasks. Experiment 3: Response time facilitation can be observed without learning Participants in Experiments 1 & 2 showed no reliable response time facilitation (Figs. 5b, 6e). Additionally, none of the conditions showed any systematic speeding or consistent facilitation across Target Detection (TD) trials (Supp. Fig. 5). This secondary observation suggests that (a) there was no increased facilitation during the TD task itself (i.e. learning during test) and (b) participants weren’t demonstrating knowledge obtained during the exposure phase, since the facilitation was consistently around zero milliseconds. Response time facilitation did not correlate with performance in either of the direct measures in any of the 9 conditions (Supp. Fig. 4). This 80 lack of facilitation even includes participants in the explicit “Feedback Training” condition, where participants performed near ceiling on the 2AFC and Creation tasks. Could the TD task be measuring entirely distinct aspects of statistical learning from the direct measures? The stream position effect Experiments 1 & 2 revealed scant evidence of implicit knowledge in response time (RT) facilitation from the TD task (Figs. 5b, 6e). We were surprised to observe no evidence of learning, given the many previously reported positive facilitation effects using similar paradigms (e.g. Turk-Browne et al., 2005; Kim et al., 2009; Campbell et al., 2012; Musz et al., 2015). In reviewing these paradigms, the most salient difference between our paradigm and previous versions was our counterbalancing procedure for ensuring targets appeared equally in all valid TD stream positions (see “Target Detection Task” in Exp. 1 Methods). This design decision was made to ensure there was no relationship between the dependent variable of interest (triplet position) and the position of targets within each TD trial (stream position). We counterbalanced the TD trials such that each image (regardless of triplet position) occurred equally often as all other images in each stream position (position 4, 5, 6, 7 or 8) (Fig. 8a). However, in previous experiments using a TD task, this was not the case: each test stream would begin with an image from triplet position 1, so that later items in a triplet (i.e. triplet positions 2 or 3) would occur later in the test stream (Fig. 8b, “Traditional TD”). Thus, triplet position and stream position were confounded in prior studies, but de-confounded in our design. 81 This confounding of triplet and stream position is only an issue if there is a general speeding effect by stream position. This means that experimenters need to ensure that response times aren’t just decreasing as a function of stream position. Therefore, we ran a post-hoc analysis of response times by TD stream position and found that participants were consistently faster at detecting the target when the target was presented later. Finding that participants detect images later in the stream faster is unsurprising. Participants are expected to learn the task parameters given enough time performing the task. Therefore, it is probable that participants detecting targets 72 times might (implicitly or explicitly) come to realize that targets only occur in stream positions 4-8 (see “Target Detection Task” above). Note that in some TD paradigms, the number of trials is in the hundreds (e.g. Otsuka et al., 2016). Thus, we might expect response times to decrease as a function of stream position, given the increasing certainty of a target with every non-target shown. This is exactly what we observed (Fig. 7a). The existence of a confounding stream position speeding effect does not rule out the possibility that some genuine learning is being assessed in the TD task. Therefore, we ran the same analysis, but further divided the response times by stream position and triplet position (Fig. 7b). This allowed us to examine if the triplet third positions were reliably faster than the triplet first positions, which would indicate that the more predictable the image, the more rapidly it was detected. However, this is not what we observed (Fig. 7b). Instead, we saw similar response times across triplet positions at each stream position. Thus, after the stream position effect was removed, there was no residual evidence of triplet learning in the response time data. 82 Figure 7: RTs across Experiments 1 & 2 by stream position in the Target Detection Task. (a) Response times by TD stream position. Sloped lines indicate the line of best fit. (b) RTs by stream position further divided by triplet position of the target. Vertical line segments indicate 95% confidence intervals. Filled circles indicate means. 83 Motivation & Preview If stream position and triplet position were confounded in previous target detection paradigms, it is important to test whether response time facilitation effects provide evidence for statistical learning. Specifically, we must test whether we can observe response time facilitation effects in the absence of any statistical structure. It is crucial to determine the validity of the standard indirect measure of visual statistical learning because one of the strongest claims about statistical learning (as a distinctive learning process) is that it occurs implicitly and does not always result in explicit knowledge. If the standard measure is not accurately assessing learning, then much of the evidence for implicit learning must be reconsidered. Given that: (i) TD stream position affects response time (Fig. 7), and (ii) prior experiments confounded stream position and triplet position in TD tasks (e.g. Kim et al., 2009), we set out to directly test whether the stream position confound could explain the response time facilitation reported in prior studies. We hypothesized that we could elicit RT facilitation by making a small change to the TD task, while preserving the same stimuli and initial exposure phase from Experiment 1 (“60 Repetitions”). To accomplish this, we intentionally introduced the same confound between stream position and triplet position that is present in traditional TD tasks. We discovered that the strongest evidence for implicit visual statistical learning (VSL) might suffer from a design confound and that previous reports of learning via this indirect measure (response time facilitation) could be explained by a confound between triplet position and test-stream position. When using our counterbalanced design, there was no evidence of VSL from response 84 time measures (Exps. 1-2), and when we introduced this confound (Exp. 3, “Traditional TD”), the ostensible signature of learning, response time facilitation, emerged. Critically, when we used the same targets, in the same target stream positions, but removed all statistical structure, we still observed the ostensible signature of learning (Exp. 3, “Scrambled TD”). Method Participants. Participants (n=240, ages 18-74; Suppl. Fig. 1) performed visual sequence learning in an identical manner to Experiment 1. We analyzed data from 120 participants in each of two conditions. Data were collected online using Amazon’s Mechanical Turk platform, coordinated using PsiTurk (T. M. Gureckis et al., 2016). Participant compensation was $4.40 (Suppl. Table 1). To ensure high data quality, we limited participation to “US only” participants who had completed at least 99% of their prior Amazon tasks successfully. Exclusion Criteria. Identical to Experiment 1. In addition, participants were prevented from participating if their Amazon Mechanical Turk ID had participated in a previous condition or pilot of this experiment. 478 subjects participated, with 180 subjects excluded and 58 held out to ensure that we had counterbalanced datasets in each condition (Supp. Table 1). Stimuli. Identical to Experiment 1 (Fig. 4a). 85 Procedure Identical to Experiment 1, “60 Repetitions”, except for the TD task changes described below. Traditional TD Task. We modified the TD task to better match the traditional testing method (e.g. Kim et al., 2009; Otsuka et al., 2016). Thus, instead of having variable length trials (Fig. 8a), where the target could appear in multiple stream positions, each of the 72 trials consisted of exactly 12 elements (i.e. all four triplets on each trial; Fig. 8b, “Traditional TD”). Following previous work (e.g. Turk-Browne et al., 2005), targets never appeared in the first three or last three positions of the stream, and so there were 6 stream positions where the target could occur. Importantly, each trial began with a first triplet position image (e.g. C1 of the triplet C1, C2, C3) and ended with a third triplet position image (e.g. C3). Crucially, this meant that targets were no longer balanced by stream position: first triplet position stimuli could only appear in stream positions 4 or 7, second triplet position images could only appear in stream positions 5 or 8, and third triplet position images could only appear in stream positions 6 or 9. Scrambled TD Task. The stream position of targets in the TD task were maintained from the “Traditional TD” condition, but the stream positions of the other 11 images were randomized (Fig. 8b, “Scrambled TD”). This procedure eliminated the triplet structure of the exposure phase from the TD task. Thus, statistical learning during initial exposure could not lead to response time speeding of later items in a triplet. All other phases of the experiment were identical to the “Traditional TD” condition (and Exp. 1). 86 Figure 8: Target Detection Task Paradigms. (a) TD task used in Experiments 1 & 2. C2 & C3 indicate trial-specific targets. Note that the stream length dictates the triplet position of the starting element. (b) TD paradigm used in Experiment 3. "Traditional TD" indicates the standard presentation of four triplets with C2 as the example, trial-specific target. “Scrambled TD” uses the same stream position of the trial-specific target, but removes all triplet structure. Subscripts (1, 2, 3) indicate triplet position, while letters (A, B, C, D) indicate triplet identity. Results Response times decreased as a function of stream position Just as in Exps. 1-2 (Fig. 7), we observed that response times decreased as targets appeared later in the TD stream (Fig. 9a). This RT speeding was approximately linear, decreasing by ~15ms per stream position. The RT speeding effect of stream position was observed regardless of triplet position (Fig. 9b). 87 Response time facilitation occurred regardless of statistical structure As hypothesized by our stream position response time speeding theory, we observed a significant difference in RTs by triplet position (Traditional TD: H(2)=82.3, p<0.001, η2=0.41; Scrambled TD: H(2)=32.1 p<0.001, η2=0.22), as well as a monotonic decrease in response times by triplet position (Fig. 9c; Table 3). Response time facilitation between triplet positions 1 & 3 was also highly significant (Fig. 9d; Table 3). Weak or no learning observed in direct measures when using Jiggle cover task Participants performed at chance levels in the 2AFC when assigned to the “Scrambled TD” condition and were just above chance in the “Traditional TD” condition (Fig. 9e; Table 3). Evidence was slightly stronger in the Creation task, but equivalent to Experiment 1 (Figs. 9f-g; Tables 2, 5). The correlation between direct measures was only significant in the “Traditional TD” condition (rs=0.305, p<0.001, Fig. 9h), though the slope was positive in “Scrambled TD” (as with all other conditions; rs =0.13). Table 5: Results from Experiment 3 (TD task comparison) Traditional TD Scrambled TD RT differences across triplet positions H(2)=82.3, p<0.001, η2=0.41 H(2)=32.1, p<0.001, η2=0.22 Mean RTs by triplet position (ms) (1st) 448.2 (2nd) 432.0 (3rd) 419.9 (1st) 449.9 (2nd) 437.3 (3rd) 430.4 Mean RTF1-3 (ms) + 95% CI 31.8, p<0.001; [31.1, 32.6] 22.2, p<0.001; [21.6, 22.9] Mean % correct (2AFC) + 95% CI 51.9, p<0.01; [51.7, 52.1] 48.9; [48.7, 49.1] 88 Mean triplets created 0.24, p<0.001; [0.23, 0.25] 0.15, p<0.001; [0.14, 0.16] Spearman rho (2AFC & Creation) 0.31, p<0.001 0.13 Pearson r (RTF1-3 & 2AFC) 0.028 0.096 Spearman rho (RTF1-3 & Creation) 0.17 0.035 Notes. H statistics are calculated with Kruskal-Wallis H tests between triplet position (see “Effect Size Analyses”). Eta squared estimate (η2) is calculated following Equation 1. No relationship between direct and indirect measures As would be expected with a metric that does not necessarily reflect learning (RTF1-3), and consistent with findings from Experiments 1-2, there was no relationship between performance in the indirect and direct measures (Supp. Fig. 4). 89 Figure 9: Results from Experiment 3. (a-d) Results from Target Detection (TD) Task. (e-h) Results from direct measures. (a) Response times by TD stream position of the target. (b) Same data as (a), but colored to show the interaction with triplet position. (c) Response times by triplet position. (d) Response time facilitation between first and third triplet positions. (e) 2AFC accuracy. (f) Creation task accuracy. (g) Number of participants creating each possible number of triplets (from 0-4). (h) Relationship between performance on the direct measures (2AFC & Creation). Dots indicate mean performance. Vertical line segments indicate 95% confidence intervals. Dashed gray lines indicate chance performance. 90 Post-hoc Analyses Learning During Test. Response time facilitation did not improve during the TD task in Exps. 1 & 2 (Suppl. Fig. 5), but did the more structured, “Traditional TD” task lead to learning during the TD task itself? Consistent with the notion that some learning occurred during the TD task, the response time facilitation increased as a function of TD task quartile, but only in the “Traditional TD” condition (Traditional RTF1-3 quartile difference H=10.3, p=0.016; Scrambled RTF1-3 quartile difference H=5.5; Supp. Fig. 6). However, comparing the RT facilitation effects between conditions revealed no significant difference (W=3073, MPRBC=-0.13). Therefore, to the extent that “real” learning occurred in the “Traditional TD” condition, it must have occurred during the TD task and not the exposure phase, since exposure was matched between conditions. If learning during test did occur, it could have been a form of procedural learning, perhaps analogous to serial reaction time (SRT) learning, which is known to be basal-ganglia dependent (Vakil et al., 2000) and is thought to be qualitatively distinct from standard statistical learning (Siegelman & Frost, 2015). These analyses indicate that some forms of learning may have occurred when the statistical structure was correlated with motor behavior, and this should inform the design of future VSL tasks. 91 Discussion Our findings might explain why conflicting evidence is observed regarding implicit and explicit processes in visual statistical learning, and suggests a reconsideration of the implicitness of visual statistical learning (Kim et al., 2009; Bertels et al., 2012). We found that a common indirect measure of visual statistical learning (i.e. the Target Detection task) was subject to a subtle interaction of triplet and stream position in response time facilitation, thought to reflect implicit VSL. This interaction was sufficient to provide evidence of learning, despite no learning occurring (Figs. 9c-d). When we employed careful counterbalancing of the Target Detection task, with an otherwise equivalent paradigm (Exp. 1, “60 Repetitions”), we found no evidence of VSL (Figs. 5a-b). Stream position was sufficient to explain apparent response time facilitation Our data suggest that stream position was the primary driver of RT facilitation. For example, in a traditional TD task, targets at stream position 7 are always a first triplet position, and so should be less predictable than targets at stream position 6 (always a third triplet position); nonetheless, targets at position 7 were detected more quickly than targets at position 6. This RT speeding by stream position occurred despite the fact that images in a third triplet position were completely predicted by both of the preceding images. Thus, although we observed robust RT facilitation, it was likely driven by the stream position effect, rather than learning of triplet sequences. The covariance of stream position and triplet position gave the appearance of RT facilitation between triplet positions, when in fact the key driver is the stream position effect (Figs. 9a-b; Table 3). 92 Participants likely learned task structure to facilitate response times Participants learned the structure of the TD task even in the absence of any sequential regularities in the visual stream (“Scrambled TD”, Fig. 9a-d; Suppl. Fig. 6). The Scrambled TD condition, which contained no statistical regularities, showed no statistically significant change in RT facilitation across quartiles (H(3)=5.5; Supp. Fig. 6), but if this approximately 22ms change from the first to later quartiles is meaningful, it would suggest that participant response times are faster after learning about task features other than triplet regularities: for example, by learning that targets can only occur in the 4th through 8th positions in a TD trial. The post-hoc analyses suggest that RT facilitation in the traditional TD task (Fig. 9d; Suppl. Fig. 6) could arise from actual triplet learning, but in this case arose primarily from a mixture of two sources: (i) the stream position effect and (ii) learning of task structure during test, rather than during initial exposure. More explicit forms of knowledge may also arise from engagement with indirect tests like the TD task, which contain triplet regularities. In the “0 Repetitions” condition of Experiment 1, there was no initial exposure phase, and yet 2AFC and Creation performance were above chance performance (2AFC mean=51.7, 95% CI=[51.5, 51.9], p=0.018; Creation mean=0.150, 95% CI=[0.142, 0.158], p<0.001). Therefore, learning could only have occurred during the TD task and was reflected in the more explicit measures of knowledge. Performance on direct measures were equivalent to Experiment 1 Evidence of learning using direct measures was equivalent to Experiment 1 (i.e. near chance; Supp. Table 3). This comparison is most relevant as the cover task was matched between Experiments 1 93 & 3. Thus, any small benefit observed in the “Traditional TD” condition was not sufficient to demonstrate significantly higher performance on the direct measures. Additionally, this serves as a conceptual replication of the “60 Repetitions” condition, as the only difference between the two TD methodology conditions and “60 Repetitions” from Experiment 1 was the TD task. Application of our findings to Target Detection tasks with longer streams In designs where streams are lengthened to 20 or more items, response times do not monotonically decrease across all stream positions (Batterink, 2017). In such designs, our preliminary analyses (data not shown) suggest that a variant of the stream position effect is still present: response times decrease monotonically with the number of images after the last target presentation. However, the stream positions and triplet positions are not perfectly confounded (only on the first target presentation) in a longer TD stream implementation: thus, for auditory tasks with long test-streams (e.g. Batterink, 2017) triplet position effects cannot be wholly explained by a confounding effect of stream position. Experiment 3 Interim Summary As hypothesized, we observed response time facilitation in the TD task, even when the ordering of items in the test stream was randomized (“Scrambled TD”). Thus, it appears that the stream position effect is a sufficient explanation for response time facilitation, producing facilitation even when learning the triplet regularities could never facilitate the responses (because they were never present in the first place). In combination with our other results, this suggests that prior 94 experiments measuring visual statistical learning in a single target TD task cannot be taken as evidence of implicit statistical learning. Instead, the regularities in the exposure phase appear to have no effect on response time facilitation (Figs. 5a, 6b). Given that the only the direct measures revealed consistent learning effects across Exps. 1-3, how does explicit knowledge correlate with performance? Explicit knowledge was evident in participants who exhibited the best learning In Experiments 1 & 2, we observed no convincing evidence of response time facilitation in the Target Detection task. In Exp. 3, we showed that the gold standard for indirect measurement of implicit knowledge (i.e. response time facilitation) could arise spuriously, when the stream position and triplet position were confounded. As such, the only convincing evidence of learning came from the direct measures, which ostensibly require more explicit knowledge. Specifically, the best-performing participants on the forced choice test were often those who could also “recreate” the stimuli (an indication of more explicit knowledge; Figs. 5f, 6c, 9h). Is it the case that beyond the direct measures potentially requiring explicit knowledge, that participants had the ability to verbalize their knowledge? If so, it would be direct evidence against the “implicit” nature of the resultant knowledge from visual statistical learning. To assess what knowledge participants had access to, either regularities or of the individual stimuli, we analyzed responses from two phases of the experiment. First, we analyzed data 95 collected before participants knew the exposure phase contained any regularities (prior to the 2AFC task) where participants were asked if they noticed a pattern. Second, we analyzed data from a post-experiment questionnaire where participants were asked whether they used labels for any of the images. Then we analyzed the relationship between their and our measures of visual statistical learning Method Patterns Seen. After participants completed the TD task, we asked participants if they “noticed any patterns in the stream of images.” If so, they were asked to “briefly describe” the pattern. Responses to this question ranged from: describing the triplet structure, mentioning the frequency of the targets in the TD task, or simply the perceptual qualities of the images. We did not subdivide the responses on quality, so responses were coded either as “yes” or “no”. Full responses can be found on OSF. Labels Used. After the experiment concluded, participants were asked if they had “named” some or all of the images presented during the experiment. If so, they were asked to provide the labels used. For those who responded “yes”, the number of labels varied from 1-12. Calculating dissimilarity of responses. It could be that participants who answer “yes” to one question are more likely to answer “yes” to another. Conversely, participants who may be eager 96 to finish the experiment might be more likely to answer “no” to all questions. Therefore, we converted each response (Y/N) for each question (Patterns Seen / Labels Used) and converted them to binary values (1 = Yes, 0 = No). We then computed the Jaccard distance between the vectors of responses for each question across participants using the distance package in SciPy 1.3.2. The Jaccard distance varies from 0-1 and provides a metric for dissimilarity between responses (where higher values are more dissimilar). Assessing differences between groups. As the sample sizes differed between groups as a result of fewer participants noticing patterns than those who did notice patterns or the number of people using labels or not, we analyzed differences between groups non-parametrically, using Mann-Whitney U tests as implemented in the Python package Pingouin 0.3.9 (Vallet, 2018). We used the same method for comparing differences within a condition, but applied a Bonferroni correction as implemented in statsmodels 0.10.2 (Seabold & Perktold, 2010) to mitigate the effects of multiple comparisons across conditions. Results Despite not being able to verbalize the triplet structure, many participants noticed repeating patterns Out of 1,080 participants where participants were naïve to the four triplets (i.e. excluding “Attend Triplets” and “Feedback Training”), 206 participants endorsed the statement that they had noticed a pattern, when probed (Suppl. Table 3). Approximately 40% of the “noticers” reported that they 97 noticed “repeated patterns”, or “order”, or “sequences”, or “pairs”, and/or “threes” (n=81). However, only 3 participants were able to verbalize that there were four triplets (all in the “Attend Relationships” condition). Noticing a pattern was not associated with improved performance in the Jiggle Task conditions, but was associated with improved performance in intentional and temporally-focused cover tasks Across all Experiments 1-3, there was a significant effect of noticing a pattern on performance in the direct measures (Pattern Seers: 2AFC mean=65.8%, Creation mean=0.97; Pattern Unseers: 2AFC mean=54.9%, Creation mean=0.40; 2AFC U=167427, p<0.001; Creation U=157851, p<0.001). However, participants in the conditions using the Jiggle cover task performed equivalently and at chance on the direct measures regardless of whether they reported seeing a pattern or not (“0 Repetitions”, “60 Repetitions”, “96 Repetitions”, “Traditional TD”, “Scrambled TD”; Fig. 10a-b; Supp. Table 3). Similarly, participants in the “Feedback Training” condition, who were already near ceiling performance, performed equivalently on all measures (Supp. Table 3). In contrast, participants who noticed a pattern in the other intentional conditions (“Attend Relations”, “Attend Triplets”) or the conditions which focused on temporal relationships and stimulus features (“2-back”), demonstrated better performance on the direct measures relative to participants who did not notice a pattern (Suppl. Table 3). Surprisingly, performance in the “1-back” condition was similar to performance in conditions employing the Jiggle Task. We discuss the implications of this in the Discussion. Finally, the “Attend Stimuli” condition showed a strong association between 98 pattern seeing and 2AFC, but did not show a significant difference in performance in the Creation task (Supp. Table 3). In general, when task instructions either required or allowed more attention to the sequential relationships between stimuli, more participants explicitly noticed patterns in the exposure stream. Moreover, the percentage of participants noticing a pattern was greater in the same conditions in which we observed a greater difference in direct measure performance (Fig. 10a; Supp. Table 3). This suggests that as more participants reported noticing a pattern, a higher percentage were successful at acting on that pattern. Noticing a pattern had no clear effect on response time facilitation None of the conditions from Exps. 1 or 2 showed a clear effect of noticing a pattern improving (or changing) performance in the TD task, as 5 conditions showed worse RTF1-3 when reporting a pattern and 4 conditions showed improved RTF1-3 (Fig. 10c; Supp. Table 3). As shown in Exp. 3, response time facilitation can occur spuriously (Fig. 9c-d) and the facilitatory effects for the “Traditional TD” and “Scrambled TD” conditions resemble mean-shifted effects from other conditions (Fig. 10c). 99 Figure 10: Performance for participants who reported either seeing or not seeing a pattern before starting the 2AFC task. (a) 2AFC performance by pattern seeing. Number of participants in each bin are shown (n=), as well as percentage of number of participants in that condition (%). The same values are true in a-c. (b) Creation performance by pattern seeing. (c) Response time facilitation between triplet first and third positions. Vertical line segments indicate 95% confidence intervals. Dashed gray lines indicate chance performance. Assigning labels to images was not associated with performance, except in cover tasks which required intent or attention to temporal relationships Out of all 1,254 participants in Experiments 1-3, about a third (n=407) indicated that they used a label during the experiment. Overall, there was a significant effect of using labels on performance in the direct measures (Labellers: 2AFC mean=62.8%, Creation mean=0.88; No Labels: 2AFC mean=54.5%, Creation mean=0.35; 2AFC U=208756, p<0.001; Creation U=207824, p<0.001). For participants in a Jiggle condition (Exps. 1 & 3), “1-back”, “Attend Stimuli”, or “Feedback Training”, performance across the three measures of learning was equivalent for labellers and non-labellers (Fig. 11; Supp. Table 3). For participants in intentional statistical learning conditions (“Attend Relations”, “Attend Triplets”) or a temporally- and stimulus-focused (“2-back”) condition, 100 assigning a label was associated with higher performance (Figs. 11a-b). Improved performance was not observed in response time facilitation data, though the “Attend Relations” condition may merit further investigation (Fig. 11c; Supp. Table 3). As with the “Patterns Seen” responses, it appears that the statistical learning conditions showing the largest difference between labellers and non-labellers also had the highest percentage of participants using labels (Supp. Table 3). A majority of participants in the “Feedback Training” condition also used labels (Fig. 11a). Figure 11: Performance across measures of learning when separating participants who assigned labels to the images and those who did not. (a) Performance on the 2AFC task. Percentages shown in parentheses represent the percentage of participants who either did or did not use labels (same values in a-c). (b) Performance on the Creation Task. (c) Performance as assessed with response time facilitation between the first and third triplet positions. Vertical line segments indicate 95% confidence intervals. Dashed gray lines indicate chance performance. Within-subject responses were highly dissimilar for “Patterns Seen” and “Labels Used” 101 Since we observed similar results in the analysis of “Patterns Seen” and the analysis of “Labels Used”, it could be that there was a high degree of overlap in participants who saw patterns and those who used labels (or who did neither). Therefore, we represented the pool of participants as two binary vectors (one for endorsement of Patterns Seen and one for Labels Used) and measured the Jaccard distance to quantify the dissimilarity between the sets of participants. The Jaccard distance between the two questions in Exps. 1-3 ranged from 0.64-0.95, with a mean distance of 0.80 (Suppl. Table 3). As “1” is most dissimilar, these values indicate that participants were not answering each question identically, but rather reflect different strategies and sources of information. Discussion Noticing a pattern, using labels, and several other post-experiment responses (e.g. “2AFC strategy used”, data not shown) indicated a positive relationship between explicit knowledge and performance. We interpret these data as implying that the nature of knowledge resultant from visual statistical learning is, at least partially, explicit. We found that a minority, but sizable subset of participants explicitly noticed patterns in the exposure phase or Target Detection task, and this is likely associated with a more explicit learning process. Similarly, participants who had both assigned labels to the images and could freely recall them demonstrated better performance on the direct measures, in some conditions. “In some conditions” is a critical qualifier, as it seems that, as we showed in Exp. 2, the cover task plays a critical role in the extent to which learning manifests. Additionally, it was the conditions 102 associated with the highest performance in which a greater percentage of participants noticed patterns or used labels. For example, in the “2-back” condition, the majority of participants assigned labels to the stimuli (>50%; Fig. 11a). Similarly, all of the conditions from Exp. 2 (except “1-back”), had >25% of participants noticing a pattern, whereas in Experiments 1 & 3, there were fewer than 20% in each condition (and only 7.5% in “60 Repetitions”) (Fig. 10a). In sum, these data suggest that, though visual statistical learning in adults may partially rely upon implicit learning processes and result in implicit knowledge, the strongest evidence of learning comes from more explicit learning and results in explicit knowledge. Why was 1-back performance so similar to the Jiggle cover task? Performance in the “1-back” condition was lower than expected from the initial design of Experiment 2, and this was also true of performance differences when comparing “pattern seers” and “labellers”. We suggest that this could be due to participants merely assessing perceptual similarity between sequential stimuli, rather than performing more holistic processing of the stimuli. Specifically, we propose two variant cover tasks that could improve performance. First, we hypothesize that a depth-of-processing story is pertinent to these data (Craik & Tulving, 1975) and that a task similar to a 1-back, but requiring more holistic processing would improve performance and increase the percentage of “pattern seers”. For example, the pre-experiment instructions could request participants imagine each image as art on a wall. Second, we hypothesize that the duration of memory maintenance is relevant. It could be that holding the 103 information in mind for a brief ISI (200ms) is insufficient to require effortful processing; participants in our task could transiently hold a limited visual feature in mind and still perform the task flawlessly. We suggest that a longer duration between repeats would necessarily improve relational learning over time. This possibility is supported by the superior performance in the 2-back condition. How implicit or explicit was the knowledge? Despite having some explicit knowledge, almost all participants were unable to explicitly explain the underlying structure (only 3 participants across Exps. 1-3; Supp. Table 3). Even though many of the responses showed that explicit knowledge was correlated with better performance (“Patterns Seen”, “Labels Used”, strategy usage), there were other questions ostensibly requiring explicit knowledge that did not demonstrate a clear pattern. For example, in the post-experiment questionnaire, participants endorsing the statement that they knew there were 4 triplets prior to the Creation task showed no systematic relationship with performance (Suppl. Table 3). This was even true for participants who created all four triplets in the Triplet Creation task. Complicating this “implicit”/”explicit” distinction further, meta-cognitive access to the knowledge was poor, and was not as good a predictor of performance relative to noticing a pattern or using labels (Supp. Table 3). Does this mean that the knowledge is implicit? Previous work has suggested this ambiguous explicit knowledge and lack of verbalizability necessitates an “implicit” label (e.g. Turk-Browne et al., 2005). We suggest instead that verbalizability of the statistical regularities is too stringent a definition for the delineation of 104 implicit/explicit memory. Many participants were able to recognize that a pattern existed (Supp. Table 3) and this explicit knowledge likely facilitated their subsequent analysis of the image stream in a more attentive and temporal way. While our paradigm does not allow for testing the triplet-by-triplet accumulation (or realization) of knowledge, previous work has suggested that statistical learning can be like a step-function, where participants very rapidly realize the structure and perform well on direct measures of learning (de Leeuw, 2016). In sum, the implicit/explicit debate remains unsettled, but our data provide critical insight into the nature of the underlying learning process and resultant knowledge. Chapter Discussion The data presented in this chapter, combined with our new analyses of data from Experiments 1-3, support the claim that visual statistical learning in adults is not always an implicit learning process and that it does not necessarily produce implicit knowledge. Explicit knowledge was especially common amongst participants who exhibited above-average learning performance (Fig. 10). Should the TD task be used to assess visual statistical learning? We employed a traditional TD task, which confounds stream and triplet position, and observed the typically reported RT facilitation effect (Fig. 9c-d). RT facilitation has been treated as evidence for implicit visual statistical learning (Kim et al., 2009), but our findings across three experiments 105 (Figs. 5a, 6c, 9c), suggest that the previous evidence of RT facilitation likely reflects an artifact of the testing paradigm. Consistent with this interpretation, there was no significant correlation between performance on the TD task and either of the direct measures (Supp. Fig. 3). It is possible to generate a response time facilitation effect, even when no structure is present in the test stream (e.g. “Scrambled TD” condition). However, when the stream position confound was present in Experiment 3, we observed a quantitatively higher response time facilitation effect for “Traditional TD” relative to the “Scrambled TD” condition. This difference was not significant, but it is intriguing that a highly structured TD task might also elicit visual statistical learning. This point is especially relevant when considering the performance on the direct measures for participants in the “0 Repetitions” condition of Experiment 1, where all learning, by definition, must have occurred during the TD task. Future work should investigate if the more highly structured “traditional” TD task would produce even better results without the interference or reduced attention from an initial exposure phase. One possibility is that the elimination of the exposure phase or a shorter exposure phase could counter-intuitively improve learning. Another possibility on why RT facilitation was at chance in Exps. 1 & 2 is that some learning was occurring in the initial exposure phase, but this learning could only be expressed when the TD task was structured as in Experiment 3 (i.e. within test streams that always began with a triplet first position). However, this theory would require that not only is the TD task exposure-dependent, but also that it measures different knowledge than the direct measures, as it could not explain the equivalent learning demonstrated in the “0 Repetitions” condition relative to the “Traditional TD” condition. Additionally, this theory discounts research suggesting that explicit knowledge can contribute to RT facilitation, but only at slower presentation rates (Arciuli et al., 2014; Bertels, 106 Destrebecqz, et al., 2015). Therefore, we suggest that the lack of RT facilitation in Experiment 1 & 2 is more likely to reflect a lack of learning or reduced ability to leverage explicit knowledge. In previous work, faster responses to later triplet positions were thought to arise from statistical learning during exposure; however, the results of this analysis suggest that learning is not necessary to explain this triplet position speeding. In the traditional TD test design, later triplet positions are always shown later in a stream relative to triplet first positions: thus, the stream position effect of faster response times to images later in the stream can explain RT facilitation effects, without any need to invoke triplet learning. Our findings serve as a reminder that minute or seemingly irrelevant features of task design can have powerful effects on human task performance. Researchers in other fields, such as in attentional capture, have noted issues arising from apparently innocuous methodological choices (DiBartolo, 2019). Unfortunately, even when methodological corrections are identified by researchers in a particular field, the importance of these methodological choices are not always widely emphasized in the literature, and the conclusions derived from flawed experiments continue to be cited. We hope our data will cause the authors employing the traditional TD task in visual statistical learning paradigms to reconsider some of their conclusions. Explicit knowledge was associated with visual statistical learning, whereas a lack of explicit knowledge was not, especially under implicit learning conditions Regardless of the mixture of implicit and explicit knowledge, the key finding of Experiment 2 (Fig. 6) was reinforced here, that large variation in performance occurred across conditions (Figs. 10- 107 11). The nature of learning and resultant knowledge seems to be associated with changes in participants’ task orientation and volition. We found evidence that explicit knowledge was associated with above chance performance, while lack of explicit knowledge was often associated with chance performance (e.g. Fig. 10). Our data is consistent with the proposal of Lynne Reder, who argues that there is no distinction between the underlying knowledge of some implicit and explicit memory tasks, stating that their performance relies upon the same memory stores in both cases (Reder et al., 2009). Any epiphenomenal differences emerge as a result of whether the task requires new associations be formed. In our experiments, the measures of testing are held constant, but the ability to learn new associations are more or less difficult during the exposure phase. An exciting possibility is that participants learning strategy rapidly shifts from an undirected one to something more explicit when they notice a pattern. This possibility is consistent with the language learning bootstrapping hypothesis (Saffran, Newport, et al., 1996; Batchelder, 2002; Abend et al., 2017), which argues that learning some language structure provides a scaffold for learning additional language structure. Many others have suggested similar theories in the statistical learning world, though there has been surprisingly little follow-up research on distinguishing the processes which give rise to noticing a regularity and then subsequent strategy use (or whether it’s the same learning process). Relatively few studies report the number of participants who recognize structure in a visual statistical learning paradigm. One exception is a study by Emberson & Rubenstein (2016). In Experiment 1 of their study, 9 of 16 participants explicitly noticed structure. Of the 16 participants, 108 half saw typical exemplars of categories (e.g. common dog) and half saw atypical exemplars of categories (e.g. uncommon bird). The 8 participants who were shown the typical images all noticed a pattern in the stream of images. Of the remaining 8, who saw atypical images, only 1 participant noticed a pattern. Additionally, participants learned the same statistical structure better when the stimuli were the typical exemplars. This suggests that there is some interaction between the choice of stimulus, explicitness, and performance during test. In Chapter 4, we investigate how choice of stimulus and prior knowledge of that stimulus affects visual statistical learning. 109 Chapter 4: Visual statistical learning differed by stimulus type The most extreme argument one could make regarding stimulus non-specificity and statistical learning is that statistical learning operates identically across all types of stimuli. While it is known that statistical learning can operate across modalities and stimulus types (Saffran, Aslin, et al., 1996; Saffran et al., 1999a; Fiser & Aslin, 2001, 2002; Christopher M. Conway & Christiansen, 2005), it is unknown how stimulus choice affects statistical learning (SL). In fact, this question is often intentionally avoided, as novel stimuli, with which participants have no familiarity, permit the investigation into “pure” statistical learning (Turk-Browne, 2012). Visual statistical learning (VSL) has been especially careful in the choice of stimuli to ensure that prior knowledge did not interfere with the study of the statistical learning process (Fiser & Aslin, 2002; Turk-Browne et al., 2005; Kim et al., 2009; Bertels et al., 2012; Bays et al., 2015; Siegelman, Bogaerts, Kronenfeld, et al., 2018; Park et al., 2018). Why? The assumption is that there would be an interaction between long-term memory of familiar stimuli and that would affect how well new associations are learned. However, this avoidance of familiar stimuli is slightly in conflict with the belief that statistical learning is an automatic learning process (i.e. “always on”), associating any relationships between stimuli which we encounter. We propose investigating this tension directly by examining how existing knowledge about the stimuli interacts with the VSL process operating on those stimuli. Familiarity with stimuli is known to play a role in how well information is learned generally, both implicitly and explicitly, and it also affects the availability of the learned information (Ziori & Dienes, 2008). One of the facilitating factors to quickly finding an object (e.g. our keys) in a cluttered visual scene is the number of times we have previously seen the object. Previous work 110 in associative learning has shown that famous faces are both remembered better than matched non-famous faces when explicitly studied, and famous faces are recognized more often than non-famous faces under conditions of incidental exposure (Klatzky & Forrest, 1984; Bird et al., 2011; Bellana et al., 2019). Implicit memory research has also shown a benefit of prior knowledge and, additively, of repetition (Schacter & McGlynn, 1989). The same prior knowledge benefit is true of working memory tasks, where words are easier to keep in mind relative to non-words (Perrachione et al., 2017) and non-words with high “word-likeness” are remembered better than low “word-likeness” words (Gathercole et al., 1991). Finally, experts are known to have better learning and memory in their field of expertise relative to novices (Ericsson et al., 1993). Collectively, these data suggest two things: (1) that stimuli are associated more easily when we have more experience with the stimuli and (2) the knowledge you bring into an experiment modulates learning. Familiarity with stimuli appears to modulate statistical learning. Typical images from a category (e.g. birds) are easier to associate with typical images from another category (e.g. dogs) relative to atypical images from the same categories (Emberson & Rubinstein, 2016). Similarly, studies using the standard SL paradigm seem to reflect better learning (e.g. in 2AFC performance) when using familiar stimuli relative to novel stimuli (e.g. Otsuka et al., 2013). Work in auditory statistical learning has shown that prior experience with a language can both interfere with learning (Finn & Hudson Kam, 2008) or facilitate learning (Siegelman, Bogaerts, Elazar, et al., 2018), depending upon whether the information to-be-learned comports with the information known. However, given the variety of cover tasks that have been used to expose participants to regularities in SL paradigms, differences in stimulus choice have gone largely uninvestigated. How is VSL affected by familiarity with the stimuli to-be-associated? 111 It seems that: (a) prior knowledge facilitates association building (as long as the association is not incongruent), (b) this benefit occurs regardless of whether the learning is intentional or incidental, and (c) the possible benefits of prior knowledge to VSL are largely unexplored. Why might visual statistical learning be affected by stimulus choice? As mentioned previously, the learning mechanisms underlying VSL are difficult to investigate due to the conceptual ambiguity in the psychological construct. However, to the extent that statistical learning shares at least some learning mechanisms with traditionally studied associative learning and memory processes (P. J. Reber, 2013; Thiessen, 2017), manipulations that benefit associative learning should also benefit statistical learning. Here we highlight several areas of research in relational processing that are candidates for influencing VSL. We approach each of the areas under the hypothesis that VSL occurs more rapidly when the constituent elements to be associated are more familiar. Specifically, we discuss perceptual fluency, depth of processing, and prior knowledge. They are separated to illustrate how they have been characterized in the field, despite potentially referring to shared underlying learning processes and neural mechanisms. For example, it seems likely that prior knowledge and perceptual fluency are highly related, but not completely overlapping concepts. Similarly, it seems probable that hierarchical processing is highly compatible with the depth of processing theory. 112 Perceptual fluency One possibility is that associability -- how easily a particular stimulus is paired with another stimulus -- increases because of perceptual fluency. Perceptual fluency is “the ability to rapidly or efficiently code and remember the stimuli that statistical learning occurs over” (Perfors & Kidd, 2020). Perfors and Kidd claim that visual statistical learning is modulated by familiarity with the stimuli, even when controlling for stimulus complexity (Perfors & Kidd, 2020). In Experiment 3 of their 2020 study, Perfors & Kidd demonstrate that native Chinese speakers process Chinese characters more fluently than native English speakers. This is unsurprising. However, they also showed that the native Chinese speakers learned an embedded statistical structure better than the native English speakers, but only when the stimuli were Chinese characters (not novel, matched symbols). Finally, they found a strong correlation between their measure of perceptual fluency and visual statistical learning. Perceptual fluency is also known to influence participants’ ability to act upon knowledge learned in artificial grammar tasks (Chang & Knowlton, 2004), which are often considered to reflect the same learning processes as statistical learning (P. Perruchet & Pacton, 2006). Depth of processing Depth (or levels) of processing (Craik & Lockhart, 1972; Craik & Tulving, 1975; Craik, 2002, 2020) is usually considered from a task perspective (e.g. counting letters in a word vs using it in a sentence). In Craik & Tulving (1975), this was shown by having participants attend to different aspects of written words. For example, participants might have to analyze the font of a word 113 (shallow processing), or whether it rhymes with another word (intermediate processing), or whether it is semantically related to another concept (deep processing). The depth of processing view is consistent with the belief that there is no distinct “memory encoding process”, but rather encoding is a byproduct of perception and comprehension (Craik, 2020). This view is consistent with the evidence we collected in Exps. 1-3 (and especially Exp. 2). In addition to the depth of processing, the potential depth of processing may also affect what is learned. For example, a word can be assessed on the font, or rhyming, or category membership, but a non-word can only be assessed on the first two. If stimuli lack the potential to be processed as deeply as otherwise similar stimuli, then, depending upon the task, learning might be more difficult. Under this assumption, we might consider famous faces to have a deeper processing potential than matched non-famous faces, and thus better VSL capabilities. Hierarchical information processing The mammalian cerebral cortex is organized as a functional hierarchy. The lowest levels of this hierarchy are located in primary sensory and motor cortices and higher levels are reached at increasing synaptic distance from the periphery (Fuster, 1997; Mesulam, 1998). Increasing selectivity for feature configurations is observed further up the cortical hierarchy (K. Grill-Spector & Malach, 2004; Rauschecker & Scott, 2009) and this observation is reflected by the fact that simple line orientation is represented in primary visual cortices (Hubel & Wiesel, 1962), whereas configural details of a face are found in not only occipital, but also temporal cortices (Kanwisher, 2010). In response to time-varying stimuli, recordings further up the hierarchy reveal longer 114 temporal receptive windows and increasing selectivity for coherent temporal structures (Xu et al., 2005; U. Hasson et al., 2008; Lerner et al., 2011; Honey et al., 2012). Overall, circuits at low levels are affected by local spatiotemporal properties such as the pitch of sounds and the orientation of image patches, while higher order circuits are most affected by more complex and spatiotemporally distributed properties, such as those that determine semantics, animacy, and object-scene relationships. Processing spatiotemporal information hierarchically suggests that, when comparing across a given task, the choice of stimulus will determine the hierarchical level of processing (Himberger et al., 2018). Concretely, we might expect that matched cover tasks will result in better VSL when the stimuli used are faces or common objects as opposed to oriented gratings or connected line segments. This suggestion is consistent with the depth-of-processing theory and provides a more concrete neural implementation. Prior knowledge The concept of “prior knowledge” includes both implicit and explicit knowledge. Both can influence our automatic and deliberative processing. For example, deviant sounds in an oddball paradigm (R. Näätänen et al., 1978; May & Tiitinen, 2010; Kremláček et al., 2016) cause a larger response when the sounds are incongruent with a participant’s first language, but not in an unlearned language (Risto Näätänen et al., 1997). Similarly, prior knowledge can interfere with incidental (but not intentional) learning (Gebhart et al., 2009). More explicit, semantic knowledge can facilitate association formation (Kole & Healy, 2007), provide scaffolding for future learning (Scaffolding Student Learning, 1997), and support transfer learning (Yurovsky et al., 2012). Prior 115 knowledge (in the form of repeated presentations) can also benefit more traditional research paradigms like list learning (Underwood, 1969). The interaction between implicit and explicit knowledge contributions to learning is difficult to assess. It is likely that semantic knowledge can automatically and implicitly activate neural circuits (Kutas & Federmeier, 2011). Conversely, implicit memory can affect explicit actions. For example, priming experiments on implicit memory have demonstrated that prior knowledge (both pre-experimental and from within-experiment) improved free association and cued recall (Schacter & McGlynn, 1989). Additionally, this study showed that a lack of prior knowledge resulted in reduced learning, highlighting that stimulus choice has a large effect on subsequent memory. Finally, research into schema instantiation, accommodation, and assimilation (Gilboa & Marlatte, 2017), suggests that the underlying neural systems processing familiar information vs novel information is categorically different and potentially graded (van Kesteren et al., 2013). This suggests that the degree of prior knowledge could affect the degree of interference or facilitation in new association building (Kesteren et al., 2018). Faces as a manipulation of stimulus familiarity We chose to use famous faces and matched non-famous faces as our stimuli. The ideal manipulation of prior knowledge would be a within-subject, parametric manipulation of the amount of stimulus exposure, beginning with a set of entirely novel stimuli. However, short-term exposure to stimuli over the course of an hour will not elicit cortical traces that are comparable to those acquired from years of real-world exposure. Thus, our approach capitalizes on the fact that 116 we have incredible amounts of real-world exposure to some stimuli. Hence, we chose perceptually similar faces that varied dramatically in prior exposure. People learn associations presented with famous faces more easily than with non-famous faces (Reder et al., 2013; Liu et al., 2017; Bein et al., 2019; Bellana et al., 2019). For example, people are better at recognizing houses paired with famous, rather than non-famous, faces (where “pairing” requires imagining the person in front of the house; Liu et al., 2017). This phenomenon seems intuitive when you consider the ease of associating a familiar person with a name or occupation, as opposed to associating a name or occupation with someone new. Schwartz and Yovel investigated this phenomenon -- of better associating familiar faces with names & occupations relative to unfamiliar faces -- and validated experimentally that familiar faces do provide an associative benefit (Schwartz & Yovel, 2016). These examples highlight how associations with famous or familiar faces are better learned than faces with which we have less experience and, more generally, how associations are influenced by the number of times they are experienced in our environment. If unfamiliar faces are processed in categorically different ways from familiar faces (Megreya & Burton, 2006; Tacikowski et al., 2011), is the same true of associations? Some neuroimaging work suggests that the answer is yes (Bar et al., 2008). Moshe Bar and colleagues found that famous faces activated the parahippocampal cortex (ostensibly processing location for contextual associations), unlike unfamiliar faces. Faces also appear to be processed hierarchically, with perceptual details processed in lower order regions and more identity-specific and social information represented in higher order regions (Castello et al., 2017). The neural representation of a familiar or unfamiliar face is categorically different. Familiar faces are so described because 117 the neural representation is robust and the ability to recall their name is much more efficient (Castello et al., 2017). We suggest that famous and non-famous faces offer the ideal stimuli for assessing visual statistical learning differences resulting from prior knowledge. In Experiment 4, we investigate if famous faces exhibit improved visual statistical learning relative to non-famous faces. Studies on associative learning suggest that pairing a known face with a novel face leads to faster decision making than two paired novel faces (Bein et al., 2019). Is the incidental visual statistical learning of famous faces also superior to non-famous faces? Experiment 4: Visual statistical learning was weaker when seeing abstract fractals relative to faces, but famous faces were not learned better than non-famous faces We designed Experiment 4 to test whether greater prior exposure to specific stimuli increased the rate at which those stimuli were associated. Previous work on perceptual fluency (Perfors & Kidd, 2020) suggests that stimuli are more efficiently processed after repeated exposure and that may correlate with visual statistical learning. Similarly, learning a pairing between a famous face and a non-famous face appears to be more rapid than pairing two non-famous faces (Bein et al., 2019). We built upon these findings to test whether associations between stimuli that have been seen more often will be learned more rapidly than between novel stimuli. We chose the “1-back” cover task as we felt it could serve as a good baseline for the hypothesized improved learning using faces. The use of the “1-back” cover task allows us to compare the effects 118 of stimulus choice while maintaining all other experimental parameters and providing an attentional check during the initial exposure phase. Given that we observed no convincing evidence of learning in the TD task in Exps. 1-3, we chose not to use that measure in Exp. 4. We predicted that performance on the two direct measures of learning (2AFC and Creation) would be better for both famous and non-famous faces relative to the abstract fractals from the “1-back” condition (Figs. 4a, 6a-b). Given that prior knowledge is largely beneficial for associative learning, implicit learning, and working memory, we hypothesized that famous faces would exhibit a learning benefit over the non-famous faces. We also hypothesized that the two direct measures would be significantly correlated regardless of stimulus type. Method Participants. Participants (n=120, ages 18-69; Suppl. Fig. 1) performed visual sequence learning in an identical manner to Experiment 2 (“1-back” condition). We analyzed data from 60 participants in each of two conditions. Since fewer participants were in each condition relative to the “1-back” condition of Experiment 2 (n=60 vs n=120), we used the first 60 stimulus sets from the “1-back” condition (see “Stimuli” in Methods). Data were collected online using Amazon’s Mechanical Turk platform, coordinated using PsiTurk 3.0.0 (T. M. Gureckis et al., 2016). Participant compensation was $4.00 (Suppl. Table 1). To ensure high data quality, we limited participation to “US only” participants who had completed at least 99% of their prior Amazon tasks successfully. 119 Exclusion Criteria. Identical to Experiment 2 (“1-back” condition). In addition, participants were prevented from participating if their Amazon Mechanical Turk ID had participated in a previous condition or pilot of this experiment. 153 subjects participated, with 12 subjects excluded and 21 held out to ensure that we had counterbalanced datasets in each condition (Supp. Table 1). Using pre-registered exclusion criteria (Supp. Table 1), we excluded participants for: changing windows away from the experiment >20 times (n=2), indicating not all stimuli were seen during the exposure phase (n=1), indicating previous participation in a similar experiment (n=3), excessive keypresses (e.g. pressed the space bar repeatedly) (n=1), not making a selection in at least 95% of 2AFC test trials (n=2), or indicating at least one set of instructions were not understood or not followed (n=3). Stimuli. The first 60 stimulus sets from Experiment 2 (“1-back” condition) were used, but instead of fractals (Fig. 4a), we used non-famous faces (Fig. 12a) and famous faces (Fig. 12b; Bellana et al., 2019). The non-famous faces were yoked to the famous faces in an attempt to match visual field, age, sex, race, and attractiveness (see Bellana et al., 2019 for more details). Procedure Experimental design was identical to Experiment 2, “1-back” condition. Analyses were identical to Experiment 2, except the “Patterns Seen” analysis which was identical to Experiment 3 and the post-hoc model comparison, which is explained below. 120 Post-hoc model comparison. To test the significance of a potential interaction between face condition and whether patterns were noticed, we built several multivariate linear models using ordinary least squares regression, as implemented in statsmodels 0.10.2. The general approach was to build models with an increasing number of variables until the Akaike information criterion (AIC) & Bayesian information criterion (BIC) values increased, then we chose the model with the lowest AIC & BIC values. Results Famous and non-famous faces were learned equivalently Performance on both direct measures was quantitatively higher for the non-famous faces group relative to the famous faces group (Fig. 12c-e), but none of these differences were significant (Non-famous: 2AFC mean=63.3%, Creation mean=0.60; Famous: 2AFC mean=59.3%, Creation mean=0.53; Supp. Table 3). Faces were learned better than fractals Kruskal-Wallis H tests revealed significant differences in performance on both direct measures between fractals, non-famous faces, and famous faces (2AFC: H=8.3, p=0.016, η2=0.026; Creation: H=16.5, p<0.001, η2=0.057; Supp. Table 3). Quantitative increases in performance were observed 121 across both measures for faces relative to fractals, with Creation performance significantly better for faces relative to fractals (fractal/famous U=2905, p<0.001; fractal/non-famous U=2781, p<0.001). Only the difference between non-famous faces and fractals was significant in 2AFC performance (U = 2689, p=0.0057). Performance on the direct measures was significantly correlated As hypothesized, and as in Exps. 1-3, we observed a significant correlation between performance in the 2AFC task and in the Creation task (Fig. 12f). This finding provides additional support for the claim that these tasks assess the same underlying learning process. 122 Figure 12: Stimuli & performance from Exp. 4. (a-b) Matched (a) non-famous and (b) famous face stimuli. (c) 2AFC performance. Each dot indicates a single participant. (d) Creation performance. (e) Creation performance showing specific numbers of triplets created in each condition. (f) Relationship between 2AFC and Creation performance. Vertical line segments indicate 95% confidence intervals and black dots indicate group means (with n>4). Dashed gray lines indicate chance performance. Noticing a pattern was associated with higher performance, but only in the famous face condition Based upon our findings in Chapter 3, we ran a post-hoc analysis on performance differences between participants who noticed a pattern and those who did not. 123 As shown above, we unexpectedly observed quantitatively higher performance in the non-famous faces condition relative to the famous faces condition (Fig. 12c-d). Upon deeper investigation of the data, we made two intriguing observations. First, the percentage of participants noticing a pattern was much greater for non-famous and famous faces (28.3% & 31.7%, respectively; Fig. 13a) relative to the “1-back” (fractals) baseline condition (15.8%; Fig. 10a; Supp. Table 3). Second, we observed potential interactions between “Patterns Seen” and stimulus choice on performance on the direct measures (Fig. 13). Figure 13: Performance in Exp. 4 on the direct measures when separating performance by whether or not the participant reported noticing a pattern. (a) 2AFC performance by patterns seen. Percentages indicate the percent of participants noticing or not noticing a pattern. (b) Creation performance by patterns seen. Familiarity with stimuli may trigger different strategies or learning processes We created several models to explain 2AFC performance using statsmodels (0.10.2) and found that the model producing the lowest AIC and BIC scores (1026 & 1037, respectively) was a model that included whether the condition used famous or non-famous faces, whether the participant noticed a pattern, and the interaction between them (y=64.7 – 11.5(famous condition) – 5.0 (pattern seen) + 24.2 (famous x pattern seen)). All coefficients had p-values <=0.003, except the “pattern seen” coefficient, which was not significant (p=0.31). As this analysis was post-hoc, we 124 cannot state with confidence that this is a reliable interaction, but this finding motivates future research to test the interaction between stimulus familiarity and the type of learning. Discussion Our hypotheses met with mixed results. Although we did observe learning benefits when participants saw faces rather than fractals, we saw no difference in performance between participants who saw famous or non-famous faces (Fig. 12). In fact, we observed quantitatively higher performance in both direct measures for the non-famous faces (Supp. Fig. 3), the opposite of what we predicted. The possibility of an interaction between performance, stimulus type, and patterns seen is intriguing (Fig. 13). Multiple possibilities are explored in the Chapter Discussion below. Chapter Discussion Does prior knowledge of the stimuli facilitate statistical learning? It would be surprising if the robust behavioral and neural differences in processing of famous and nonfamous faces (Megreya & Burton, 2006; Guillaume et al., 2009; Tacikowski et al., 2011) had no impact on how visual statistical learning operates on such stimuli. However, our data, at face value, showed little difference between the famous and non-famous conditions. One possibility is that visual statistical learning is an entirely distinct process from the other types of learning which have revealed face 125 familiarity effects. More likely, as we discuss below, is that a combination of prior knowledge and task strategy effects account for the apparent equivalence of face conditions. First, it is possible that visual statistical learning was somehow impeded by the fact that the stimuli to-be-associated were famous faces. Predictive processing theories suggest that environmental expectations influence cortical activity (Kok et al., 2017). If famous faces produce some expected associations with other famous people, it might be more difficult to associate three previously unassociated, well-formed schemas that are, in some ways, incongruent (Brod et al., 2015; Gilboa & Marlatte, 2017). Famous faces, because of all their inherent associations (e.g. with emotions, relationships to other known people) might be processed in a categorically different way than stimuli lacking strong schemas. In contrast, non-famous faces might lack these pre-existing associations and, therefore, are easier to arbitrarily associate. This interpretation is consistent with two further facts. First, visual statistical learning was worse when using the fractals in a matched cover task (Exp. 2, “1-back”; Supp. Table 3). This suggests that unfamiliar stimuli are not necessarily learned robustly. Second, prior research has suggested an associative learning benefit of famous faces relative to non-famous faces (Bein et al., 2019). However, this famous face benefit was observed when a famous and a non-famous face were paired (Bein et al., 2019). It could be that because there was no schematic interference between two famous faces, that the learning benefit was observed. In our design, when all three stimuli were famous faces, the multiple, non-overlapping associations may have negated any individual stimulus’ familiarity benefit. 126 A second potential reason for the equivalence of learning across the face conditions is that neither type of stimulus was processed beyond surface-level perceptual features. Face recognition improves specifically after conceptual, rather than perceptual processing (Schwartz & Yovel, 2019). This suggests that a depth-of-processing framework (Craik & Tulving, 1975) could be relevant to face processing, and provides theoretical support for the idea that the method of engagement with the stimuli is predictive of visual statistical learning (as shown in Exp. 2). Consider two cover tasks used in visual statistical learning experiments: the 1-back task and an image categorization task. As mentioned in the Chapter 3 discussion, a 1-back task can be performed flawlessly with minimal processing of the stimuli themselves. Even within the family of 1-back tasks, ours requires minimal (200ms) maintenance of perceptual information. Compared to other 1-back tasks (e.g. Turk-Browne et al., 2005; Musz et al., 2015), which can require 800ms of maintenance and two intervening images, our task requires much less maintenance. Now consider an image categorization task (e.g. Turk-Browne et al., 2010; Vickery et al., 2018). In these experiments, the stimuli must be “deeply processed” to assess which conceptual category contains that particular stimulus; processing consistent with a depth of processing framework. It could be that no difference was observed between the faces conditions because neither stimulus type was processed beyond surface level perceptual features, despite participants knowing who the images represented when explicitly asked. This depth-of-processing interpretation is also consistent with Experiment 4 of Vickery et al. (2018), where participants incidentally familiarized to image pairs using a Jiggle task had lower recognition performance relative to participants familiarized in an image categorization task (Vickery et al., 2018, StSr condition). 127 A final possibility is that prior knowledge and cover task interact with one another. This theory finds support in the potential interaction between face condition and the “Patterns Seen” variable in predicting performance in the direct measures. If this interaction holds up in subsequent replications, it would strongly suggest that stimulus type interacted with the method of stimulus engagement, and, specifically, that visual statistical learning was only boosted when a pattern was noticed with familiar stimuli. This would also explain why “pattern noticers” in the non-famous face condition had no learning benefit (Fig. 13). We conclude this chapter by noting two possible follow-up experiments that could clarify the relationship between prior knowledge and visual statistical learning. First, following Bein et al. (2019), one could test whether mixed triplets of one famous and two novel faces are learned better than all famous or all non-famous faces. This would show that schematic processing in Experiment 4 was sufficient to inhibit a familiarity benefit otherwise observed from a single, familiar image. Second, one could test whether deeper processing of the individual stimuli might lead to not only differences between the famous and non-famous conditions, but specifically better learning in the famous faces condition. This could be accomplished by asking the participants to imagine the person represented by the current image speaking to the person previously shown. If deeper processing is possible with prior knowledge, then the same task should preferentially benefit the famous faces. Future research is necessary to disentangle the contributions of prior knowledge to the creation of new associations in the context of visual statistical learning. 128 Chapter 5: Conclusions & Open Questions As reviewed in Chapter 1, the definition of statistical learning can vary from essentially all learning (P. J. Reber, 2013), to simply the implicit detection of regularities in the environment (Arciuli, 2017). The data presented in Chapters 2-4 suggest two claims that prompt a reconsideration of what is being measured in the standard visual statistical learning (VSL) paradigm: (1) that VSL is not automatic, even provided selective attention and (2) that less implicit knowledge is produced than previously appreciated. These data suggest that much of the VSL observed in standard research paradigms can be explained without the need for invoking a new and distinct “statistical” learning mechanism. Instead, long-studied learning mechanisms could support the statistical learning phenomenon (P. J. Reber, 2013; Schapiro et al., 2016; Thiessen, 2017; Christopher M. Conway, 2020). In the following sections, I review the data and arguments from Chapters 2-4 and discuss the future implications for the conception and practice of statistical learning research. Visual Statistical Learning Characteristics Chapter 1 provided a broad overview of the field of statistical learning and suggested a cluster of defining characteristics for statistical learning: automaticity, implicitness, and stimulus-nonspecificity. In Chapter 2, we presented evidence from two experiments suggesting VSL is not automatic. The first experiment showed that VSL does not always occur, even when selective attention is deployed to the stimuli. The second experiment showed that the method of stimulus engagement during the exposure phase of the standard research paradigm strongly modulates learning. In Chapter 3, we provided evidence suggesting that VSL does not necessarily produce 129 implicit knowledge. In particular, we demonstrated that spurious evidence of VSL can be elicited when stream and triplet positions are confounded in target detection tasks. Additionally, we showed that participants who explicitly noticed patterns during the exposure phase and/or TD task were also the ones who performed well on the direct measures of learning, while there was very little learning evidence amongst participants who lacked explicit knowledge. Finally, in Chapter 4, we showed that VSL performance depends on the stimulus set: sequences of face images (both familiar and unfamiliar) were learned more rapidly than sequences of arbitrary geometric fractals. Automaticity In Chapter 2, we examined the automaticity of statistical learning by exposing participants to visual triplet sequences and comparing their performance across different exposure conditions and measures of learning (Fig. 4). When tasked with detecting image motion within a stream of images (Exp. 1), participants unexpectedly demonstrated no evidence of learning using response time facilitation as a dependent measure, and they exhibited very weak evidence of learning using two direct measures of learning (Fig. 5). Next, we assessed VSL using six different cover tasks (Exp. 2). We again observed no evidence of learning in response time facilitation, but significant differences in performance on the direct measures between conditions (Fig. 6). In Chapter 3, we conceptually replicated the “60 Repetitions” condition from Experiment 1 and again found little evidence of learning in the direct measures (Figs. 9e-g). 130 Our findings suggest that VSL may not be as automatic as commonly envisaged (Saffran et al., 1999b; Fiser & Aslin, 2002; Turk-Browne et al., 2005, 2010; Aslin & Newport, 2012; Batterink, Reber, & Paller, 2015; Batterink et al., 2019; Christopher M. Conway, 2020). In our data, we only observe learning under conditions that encouraged processing of the temporal aspects of the stimulus sequence. Thus, VSL appears to require a more active form of engagement, particularly with the temporal aspects of the visual stream. Implicit Knowledge In Chapter 3, we investigated the nature of knowledge resulting from the standard VSL research paradigm. Two main findings were presented: (1) the standard measure of implicit knowledge is confounded, and therefore the conclusions drawn from experiments using this particular implementation must be reconsidered, and (2) that more explicit knowledge is available to participants than would be expected if VSL were primarily an “implicit” process. Considerations for the Target Detection task Indirect measures of VSL, such as the target detection (TD) task, are considered to provide the clearest evidence of implicit knowledge of incidentally learned regularities. Direct measures of VSL, such as 2AFC or Creation tasks, may draw on more explicit knowledge (Turk-Browne et al., 2005; Bertels, Destrebecqz, et al., 2015; Batterink, Reber, Neville, et al., 2015). The TD paradigm has been used to investigate the dependence of VSL on attention (Turk-Browne et al., 2005; Musz et al., 2015), the permanence of knowledge acquired via VSL (Kim et al., 2009), how VSL is affected 131 in aging (Campbell et al., 2012), and the underlying neural correlates of VSL (Otsuka & Saiki, 2017). We suggest that these conclusions should be reassessed in light of the stream position speeding effect (Fig. 7) and our demonstration of VSL where none exists (Fig. 9). Previous work has suggested that VSL produces implicit and long-lasting (>24 hours) knowledge (Kim et al., 2009). However, given that stream position confounds are sufficient to produce the speeding effects from prior experiments (Fig. 9, “Scrambled TD”), these conclusions need to be revisited. The corrected, counterbalanced TD task never produced any clear evidence of VSL, even when participants had explicit knowledge of the regularities (Fig. 6d-e, “Feedback Training”). Assuming that the single-target TD task is a valid measure of implicit knowledge, future studies will need to explain why it provides no evidence of knowledge after de-confounding the stream and triplet positions. Explicit Knowledge Though VSL and performance during test can be influenced by explicit knowledge (Bertels, Boursain, et al., 2015; Bertels, Destrebecqz, et al., 2015), the core learning processes have been conceived of as arising without requiring explicit knowledge (Pierre Perruchet & Pacton, 2006). Indeed, statistical learning is often described as an unconscious (Turk-Browne et al., 2010; Turk-Browne, 2012) or incidental (Saffran et al., 1997) cognitive process. However, there have been several suggestions that explicit processes do contribute to VSL (Cleeremans, 2006; Pierre Perruchet et al., 2006; Batterink, Reber, Neville, et al., 2015; Batterink et al., 2019). 132 Our data demonstrate that explicit instruction to attend to temporal structure can lead to robust learning (e.g. Exp. 2, “Attend Triplets”), for the same design and stimuli which show little evidence of learning with a less temporal and more passive cover task (e.g. Jiggle task conditions, Exps. 1 & 3). Additionally, we suggest that encouraging relational processing (e.g. Exp. 2, “2-back”) will potentially draw attention to relationships between stimuli (Turk-Browne et al., 2005; Brady & Oliva, 2008; Arciuli & Simpson, 2012; Campbell et al., 2012; Stevens et al., 2014; Musz et al., 2015; Vickery et al., 2018), which could lead to explicit knowledge. Similarly, attributing mnemonics to the stimuli could improve relational processing. We observed the highest percentage of participants explicitly labelling stimuli in the “2-back” condition (Fig. 11). Finally, across all experiments and conditions, the direct measures of VSL were consistently positively correlated (Supp. Table 3): participants who expressed greater “familiarity” with the image regularities (assessed via 2AFC) could also explicitly create more of the sequence structures. In conditions that generated successful VSL (Fig. 6), the data could be usefully subdivided by explicit knowledge. Specifically, we found that participants who noticed patterns in the exposure phase and/or TD task demonstrated significantly better VSL than those who did not (Fig. 10). Similarly, participants who were able to explicitly verbalize labels for at least some of the stimuli demonstrated significantly better VSL than those who did not (Fig. 11). While post-experiment report does not tell us if the learning was implicit or not, it does suggest that the resultant knowledge seems to be explicit. Collectively, these data suggest that VSL paradigms generate a much greater proportion of explicit knowledge (and therefore a smaller proportion of implicit knowledge) than has been empirically appreciated or conceptualized (P. Perruchet & Pacton, 2006). 133 Stimulus Non-specificity In Chapter 4, we investigated the effects of stimulus choice and familiarity on visual statistical learning. The learning processes involved in processing any stimuli could vary according to perceptual fluency (i.e. the familiarity, modality, and complexity of stimulus sets), depth of processing (i.e. the cognitive demands of the exposure phase), prior knowledge, and the developmental stage of participants (Craik & Lockhart, 1972; Siegelman, Bogaerts, Elazar, et al., 2018; Arciuli, 2017; Arciuli & Conway, 2018; Frost et al., 2015). We originally hypothesized that famous faces would result in superior VSL relative to non-famous faces, but that is not what we observed. Previous research has suggested a learning benefit of familiarity (Otsuka et al., 2013; Emberson & Rubinstein, 2016), so it remains to be seen if the lack of a difference between face conditions (familiar vs. unfamiliar) is an effect specific to face stimuli or the degree of familiarity. The data from this chapter do not provide strong evidence for or against the broader stimulus non-specificity of visual statistical learning, but they do demonstrate that stimulus sets may change the range of strategies that participants employ, as well as their overall performance on direct measures; thus, in order to synthesize general properties of VSL from future studies, it will be helpful if those studies employ consistent stimulus sets and cover tasks. 134 Reconsidering Visual Statistical Learning SL can be considered an automatic, implicit, and stimulus-nonspecific learning process. This constellation of characteristics motivated the conception of SL as a new and distinct learning process. However, the data presented in Chapters 2-4 suggest that, at least for visual forms of SL, the tasks which we have been using to measure VSL have not reliably been measuring processes that are automatic, implicit, and stimulus-nonspecific. Below, we suggest how to move forward in light of these data. Specifically, what types of research should be conducted to better distinguish VSL from other learning processes, or to better understand how the learning processes overlap? Automaticity Statistical learning has been defined as an “automatic sensitivity to conditional probabilities” (Fiser & Aslin, 2002) and “unconscious [automatic segmentation]” (Turk-Browne et al., 2010). Experiment 1 provides large-sample evidence to suggest that the learning commonly measured in statistical learning experiments is not automatic, because we could vary learning performance from near-chance levels to high performance simply by changing the cover task. Future studies usefully investigating the boundaries of automaticity could highlight why we did not observe VSL evidence. For example, perhaps our stimulus duration or ISI were too short or too long. Alternatively, it could be the interaction of the timing, stimulus choice, and/or testing measure resulted in no evidence. As suggested by Experiment 3, some measures of VSL have previously reported evidence of learning where there was none. 135 To provide stronger support for claims of automatic regularity generalization (e.g. Marcus et al., 1999), new experimental paradigms testing automatic generalization will be necessary. For example, when category-category pairings are statistically learned (e.g. fruits – tools), can novel stimuli (e.g. blueberry – hammer) from the learned categories be processed as if they had been presented? Luo & Zhao (2018) showed such an effect by presenting pairs of city names in a 1-back VSL paradigm and found that participants were able to show transitive learning for the country pairings that contained the presented cities. Approximately 10% of the participants in their category pairing experiments (4-7) reported noticing a pattern, but could not verbalize the presented pairs. This suggests that, similar to what we observed in Experiment 3 (Fig. 10), noticing a pattern could lead to above average group effects. In the context of such results, it will be important for future studies to directly measure the interaction between associative generalization (i.e. transfer learning; Barnett & Ceci, 2002) and explicit knowledge. One final possible area of investigation about the automaticity of statistical learning is that it could be modality-dependent. Auditory processing is known to be biased towards time, whereas visual processing is biased more towards space (Himberger et al., 2018). This could lead to a reduced learning capacity of temporal regularities in the visual modality. We suggest matching stimuli across modalities by norming stimuli by perceptual fluency, then assessing if regularities are learned to the same degree with normalized stimuli across modalities. If auditory is superior, it could suggest that automatic VSL is reduced relative to ASL. 136 Implicitness As shown in Table 4, implicitness can refer to the instructions, the learning process, or the resultant knowledge. Visual statistical learning has been conceptualized to be implicit across all three. However, our data from Chapter 3 suggest that more explicit knowledge results from VSL in the standard paradigm than previously appreciated. Here we discuss several issues in the field and how they can be addressed to support the implicitness of visual statistical learning. Some SL paradigms warn participants of upcoming structure, potentially moving them from implicit instructions to explicit instructions (e.g. “Attend Triplets” condition). Consider the performance differences observed between the “Attend Triplets”, “Attend Relations”, and “Attend Stimuli” conditions, where there was a clear increase in learning as minor wording changes oriented participants towards the temporal structure of the stimuli (Fig. 6). Similarly, many auditory statistical learning paradigms instruct participants that they may become aware of patterns or that they’ll listen to an unknown “language” (Saffran, Newport, et al., 1996; McNealy et al., 2006). We propose that paradigms which allow for such explicit instruction do not satisfy definition of statistical learning, especially if statistical learning is to be distinguished from other forms of learning. It is crucial to ensure participants remain naïve to the upcoming structure when evaluating an implicit learning process. Emphasizing relational processing could preferentially induce explicit learning and/or explicit knowledge. A standard cover task in VSL is the n-back task (Exps. 2 & 4). Even though the learning is thought to be incidental and implicit, the task requires relational processing between elements in the stream. Thus, n-back tasks may draw especial attention to temporal regularities by necessity. We suggest that future VSL studies employ a detailed debriefing, using a funnel procedure that 137 begins with general questions and progresses to more specific questions, to assess the extent of explicit knowledge. Participants’ ability to verbalize the sequential regularities (which is commonly measured in SL experiments) does not seem to reliably predict task performance, but more diffuse knowledge (e.g. “did you notice a pattern?”) did predict performance (Fig. 10). While intentional learning is generally thought to be superior to unintentional learning (Eagle & Leiter, 1964), it has been suggest that explicit intent to learn could actually impair statistical learning. If this dissociation is reliable, then it could potentially be leveraged to experimentally separate the contributions of explicit and implicit learning processes in statistical learning paradigms. For example, explicit attempts to learn a motor sequence inhibited the ability to learn a sequence (Fletcher et al., 2005). Additionally, McNealy and colleagues found that stressing the starting syllable of a third of auditory triplets led to a quantitative decrease in performance relative to an equivalent, unstressed stream of syllables (McNealy et al., 2006). Similarly, Finn and colleagues found that an intentional learning strategy for phonological structure in an auditory statistical learning task resulted in chance performance, whereas a passive strategy revealed positive category learning (Finn et al., 2014). In the same study however, an intentional strategy was superior for the type of learning studied in the standard research paradigm. Other studies have observed increased statistical learning performance with explicit instruction (Yang & Li, 2012; Stevens et al., 2014; Batterink, Reber, Neville, et al., 2015). The strongest argument against the possibility that explicit intention hurts VSL, however, is the data from Experiment 2 (Fig. 6). In that experiment, participants showed much stronger VSL when they were given explicit instructions. Future work should investigate the conditions under which intentional learning strategies result 138 in reduced VSL relative to unintentional learning strategies, while equating attention to the stimuli and regularities. The incidental learning of temporal associations is not limited to statistical learning paradigms, but is a long-standing component in assessing hippocampally-dependent forms of episodic and associative learning. For example, the Temporal Contiguity Effect (TCE) is a common memory phenomenon that is exhibited when the recall of one memory triggers recall of other memories which occurred nearby in time (Kahana, 1996). The TCE is most often associated with explicit-encoding list learning paradigms, but has also been observed during incidental encoding (Healey, 2018). List learning paradigms using free recall are paradigmatic examples of hippocampally-dependent explicit learning – and if temporal associations are encoded incidentally, then one should ask whether similar processes could be operating in VSL paradigms. Thus, it is important to ask whether we might expect to observe bi-directional temporal relationships in the statistically learned sequences (i.e. not only forward triplets, but jumbled or reversed triplets). Indeed, facilitated performance for forward and backward sequences has been observed in several VSL studies and across learning measures (Otsuka et al., 2016; Otsuka & Saiki, 2017; Zhou et al., 2020), suggesting that the learning in classic list learning paradigms and VSL paradigms may have more in common than has been recognized. Healey (2018) had participants incidentally learn a list of 16 words by asking them to create a mental movie given the first word and then trying to integrate every subsequent word into their movie. This paradigm is most related to our “Attend Relationships” condition, though the Healey paradigm requires deeper processing. This approach produces a TCE (as do others that require stimulus-specific processing, such as animacy judgements). In the mental movie condition of Healey (2018), approximately 1/3 of participants 139 spontaneously thought they might receive a surprise test on the words following movie generation. This is approximately 2-3 times as many participants as in Healey’s six other matched conditions, where participants were encouraged to employ stimulus-specific and not relational processing. This suggests two things: (1) that relational processing across paradigms results in an increase in explicit knowledge and (2) participants should be debriefed to determine if they explicitly considered the relationships between stimuli. Excluding participants who explicitly considered the relationships between stimuli (as opposed to those who expected a surprise test, as were excluded in Healey (2018)) might eliminate the TCE effect, as well as VSL effects in our paradigm. The relationship between visual working memory (VWM) and visual statistical learning was investigated by Brady and colleagues (Brady et al., 2009). They suggested that VSL “compressed” stimulus representations to allow more information to be held in VWM. However, when this study was replicated, researchers found that these effects were limited to the subjects with explicit knowledge of the regularities (Ngiam et al., 2019). Additionally, these explicit knowledge participants were thought to have used an explicit strategy leveraging their long-term memory to support increased visual working memory. Future research should carefully consider how explicit knowledge interacts with VWM and whether VSL contributes uniquely to VWM capacity. Finally, the robustness of visual statistical learning should be reevaluated. One of the most convincing demonstrations of the long-lasting effects of implicitly-learned VSL (cited more than 100 times) was provided by Kim and colleagues (Kim et al., 2009). However, Experiment 3 clearly shows that an almost identical paradigm can produce ostensible “learning” effects even when there are no regularities. Recently Rawal & Teng (2021) used a similar set of paradigms to our 140 approach (Himberger et al., 2019), and confirmed that there is little evidence of implicit VSL lasting longer than a single-day session (Rawal & Tseng, 2021). To better understand the robustness of VSL effect, future studies should investigate the interaction of testing delay and implicit knowledge. Infant statistical learning does not necessarily imply implicit & non-hippocampal learning The original Saffran et al. (1996) statistical learning study generated excitement for at least two reasons (Saffran, Aslin, et al., 1996). First, the data was strong support for an innate learning mechanism that could parse statistical properties of the world. This was especially relevant in the discussion of language learning and whether infants require an innate language module to learn language. Second, and more subtly, was that the learning occurred rapidly. Infants were able to learn the statistical structure of a nonsense “language” within two minutes. Implicit in this second point is that infants do not have fully formed hippocampi (Nadel & Zola-Morgan, 1984) and therefore, the learning was unlikely to be hippocampally-dependent. Altogether, the original data suggested not only that there was an innate (or early-developing) learning mechanism for extracting the statistical regularities from the environment, but that this learning process was potentially cortically-based and domain general (Kirkham et al., 2002). Does the infant learning data support the claim that statistical learning is distinct from other learning processes? If we were to make the assumption that infants, and babies more generally, learn passively like sponges (e.g. Saffran et al., 1997), then indeed it would seem that this is a learning process quite different from what was studied in (say) list-learning experiments in adults. 141 However, it is known that 16-month olds can, in fact, form episodic memories that last for days or weeks (Bauer & Leventon, 2013). This capability is thought to be hippocampally-dependent (Ramsaran et al., 2019). These recent observations open the possibility that infants share more of their neural learning circuits with adults than was previously appreciated, especially the circuits relevant for the rapid learning tested in standard SL paradigms (i.e. memory on the order of minutes). More generally, Raz and Saxe (2020) suggest that much of the network connectivity between prefrontal cortex and other brain regions required for complex social interactions are available by 8-months. They argue this connectivity enables babies to deploy more intentional learning strategies than previously appreciated. Therefore, infants may deploy selective attention and learning goals to a greater degree than previously appreciated (Raz & Saxe, 2020), and infant and adult statistical learning may be more similar than originally envisioned (C. T. Ellis et al., 2020; Yates et al., 2021). Stimulus Non-specificity Although statistical learning researchers would likely agree that SL transpires across all stimulus types, they are unlikely to endorse the claim that it occurs equivalently regardless of familiarity, semanticity, contextual encoding similarity, or temporal proximity of the associations. As such, this particular characteristic can best be used to investigate how, if at all, VSL differs from standard associative learning in regard to the modulatory aspects mentioned. We showed in Chapter 4 that stimulus choice can improve (with face stimuli) or worsen (with fractal stimuli) VSL. The benefits of familiarity and prior knowledge are well-known in the associative learning field 142 (Liu et al., 2017; Kesteren et al., 2018; Bellana et al., 2019). How does something like the temporal contiguity effect play out in a VSL paradigm? We suggest instead of exposing participants to four triplets, researchers could use three quintets. Using a longer “word”, along with a multi-alternative forced choice test could allow researchers to assess how participants recall stimuli after VSL exposure and directly compare their performance to the TCE literature. Should we draw any distinction between two cognitive processes that form associations using the same neural circuits? For the type of rapid visual statistical learning studied using the standard research paradigm, it appears the neural circuits overlap between associative and statistical learning (Curran, 1997; Schapiro et al., 2012, 2014; Schlichting et al., 2017). Others have raised similar issues regarding shared memory processes, specifically asking the difference between statistical learning and state-dependent learning, paired-associate learning, and implicit learning (Thiessen, 2017). Many of the processes that support learning in traditional associative learning paradigms appear to be compatible with statistical learning effects, but these connections are not often emphasized in the literature. For example, there are implicit associative learning paradigms (Bellana et al., 2019) and the resultant associative knowledge can be implicit (Degonda et al., 2005). As stated in the Chapter 4 Discussion, it would be interesting to assess how triplets consisting of 1 famous and 2 non-famous faces are associated relative to all famous or all non-famous faces. Previous work has suggested that neural representations of items can become more similarly represented, but only when at least one stimulus is unfamiliar (Schapiro et al., 2012; Bein et al., 2019). 143 Integrating with other fields Many of the observations and debates in statistical learning would benefit from contextualization from the literatures on memory storage (Reder et al., 2009), schematic processing (Gilboa & Marlatte, 2017), cortical information processing and plasticity (P. J. Reber, 2013), event segmentation (Ezzyat & Davachi, 2014), and encoding vs retrieval processes (Prince et al., 2005; Davachi, 2006; Mayes et al., 2007). For example, the distinction between recognition and recollection memory has generated a vast literature (A. Yonelinas, 1999; A. P. Yonelinas et al., 1999; Mickes et al., 2009) and this work is surely relevant to understanding the pattern of relationships we observed between the TD, 2AFC, and triplet creation tasks. If we identify commonalities between statistical learning and other, more established, cognitive learning processes, this may help to integrate theories of statistical learning with these other fields. Another question is whether the rapid learning studied in a standard SL paradigm (i.e. <1 hour) is undergirded by the same learning process as more gradual SL (e.g. >3 weeks). Learning on both timescales is affected by attention, engagement, and prior knowledge (Henke, 2010; Uri Hasson et al., 2015). However, the underlying neural processes appear to be at least partially separable (Schapiro et al., 2012; P. J. Reber, 2013). In future work, we should reconsider how applicable the findings from single session studies are to understanding the gradual extraction processes that may unfold over many weeks in real-world settings. 144 Summary We found that the learning processes measured in visual statistical learning experiments are not automatic, and that the best evidence of learning is associated with explicit rather than implicit knowledge of the regularities. In addition, we found that performance in visual statistical learning tasks was sensitive to the class of stimuli that were to be associated. We did not find visual statistical learning to be an automatic process, even when attention was directed to the stimuli. Rather, it appears that some further active engagement with the stimuli is required to extract sequential regularities. We found no evidence of visual statistical learning when it was assessed via the standard indirect measure (i.e. the target detection task). We found dramatic differences in visual statistical learning depending on how participants were instructed to engage with the stimuli. Participants who recognized visual sequence regularities in a forced-choice task could also often recreate the sequences when explicitly probed, indicating their knowledge was not entirely implicit and that both measures likely reflect a singular learning process. Similarly, participants who explicitly noticed a pattern prior to test demonstrated the highest degree of learning. Statistical learning, as originally conceived, was a unique and innovative psychological construct; however, with the gradual broadening of its definition, and the proliferation of experimental paradigms under the banner of statistical learning, it is less clear how to situate this concept, or how helpful it is in organizing our understanding of learning. If the learning from the standard VSL research paradigm can be explained by other learning processes (such as procedural and 145 associative learning), then it may be parsimonious to replace or redefine the psychological construct. By the same token, it seems critical to design experiments that isolate the distinctive features of visual statistical learning. Importantly, we require designs in which (1) the regularities to be learned are temporal rather than spatial, (2) there is no motoric sequence reliably related to the learned statistical structure, and (3) that the instructions and learning are ostensibly implicit. A promising step in this direction was taken by Zhou et al. (2020), who have demonstrated differential effects of interleaved vs. blocked learning when participants were trained using an implicit, but quasi-temporal VSL paradigm vs an explicit, single-trial paired-associates paradigm. This type of cross-disciplinary approach, considering VSL in the context of learning more generally, will be crucial if we are to thoroughly characterize how humans gradually extract the regularities they encounter in the visual world. 146 Bibliography Abend, O., Kwiatkowski, T., Smith, N. J., Goldwater, S., & Steedman, M. (2017). Bootstrapping language acquisition. Cognition, 164, 116–143. https://doi.org/10.1016/j.cognition.2017.02.009 Arciuli, J. (2017). The multi-component nature of statistical learning. Phil. Trans. R. Soc. B, 372(1711), 20160058. https://doi.org/10.1098/rstb.2016.0058 Arciuli, J., & Conway, C. M. (2018). The Promise—And Challenge—Of Statistical Learning for Elucidating Atypical Language Development. Current Directions in Psychological Science, 0963721418779977. https://doi.org/10.1177/0963721418779977 Arciuli, J., & Simpson, I. C. (2012). Statistical learning is lasting and consistent over time. Neuroscience Letters, 517(2), 133–135. https://doi.org/10.1016/j.neulet.2012.04.045 Arciuli, J., Torkildsen, J. von K., Stevens, D. J., & Simpson, I. C. (2014). Statistical learning under incidental versus intentional conditions. Frontiers in Psychology, 5. https://doi.org/10.3389/fpsyg.2014.00747 Aslin, R. N. (2017). Statistical learning: A powerful mechanism that operates by mere exposure. Wiley Interdisciplinary Reviews: Cognitive Science, 8(1–2), e1373. https://doi.org/10.1002/wcs.1373 Aslin, R. N., & Newport, E. L. (2012). Statistical Learning: From Acquiring Specific Items to Forming General Rules. Current Directions in Psychological Science, 21(3), 170–176. https://doi.org/10.1177/0963721412436806 Avarguès-Weber, A., Finke, V., Nagy, M., Szabó, T., d’Amaro, D., Dyer, A. G., & Fiser, J. (2020). Different mechanisms underlie implicit visual statistical learning in honey bees and humans. Proceedings of the National Academy of Sciences, 117(41), 25923–25934. https://doi.org/10.1073/pnas.1919387117 Baddeley, A. (2012). Working Memory: Theories, Models, and Controversies. Annual Review of Psychology, 63(1), 1–29. https://doi.org/10.1146/annurev-psych-120710-100422 147 Baker, C. I., Olson, C. R., & Behrmann, M. (2004). Role of Attention and Perceptual Grouping in Visual Statistical Learning. Psychological Science, 15(7), 460–466. https://doi.org/10.1111/j.0956-7976.2004.00702.x Bar, M., Aminoff, E., & Ishai, A. (2008). Famous Faces Activate Contextual Associations in the Parahippocampal Cortex. Cerebral Cortex, 18(6), 1233–1238. https://doi.org/10.1093/cercor/bhm170 Barakat, B. K., Seitz, A. R., & Shams, L. (2013). The effect of statistical learning on internal stimulus representations: Predictable items are enhanced even when not predicted. Cognition, 129(2), 205–211. https://doi.org/10.1016/j.cognition.2013.07.003 Barnett, S. M., & Ceci, S. J. (2002). When and where do we apply what we learn?: A taxonomy for far transfer. Psychological Bulletin, 128(4), 612–637. https://doi.org/10.1037/0033-2909.128.4.612 Batchelder, E. O. (2002). Bootstrapping the lexicon: A computational model of infant speech segmentation. Cognition, 83(2), 167–206. https://doi.org/10.1016/S0010-0277(02)00002-1 Batterink, L. J. (2017). Rapid Statistical Learning Supporting Word Extraction From Continuous Speech. Psychological Science, 0956797617698226. https://doi.org/10.1177/0956797617698226 Batterink, L. J., Paller, K. A., & Reber, P. J. (2019). Understanding the Neural Bases of Implicit and Statistical Learning. Topics in Cognitive Science, 11(3), 482–503. https://doi.org/10.1111/tops.12420 Batterink, L. J., Reber, P. J., Neville, H. J., & Paller, K. A. (2015). Implicit and explicit contributions to statistical learning. Journal of Memory and Language, 83, 62–78. https://doi.org/10.1016/j.jml.2015.04.004 Batterink, L. J., Reber, P. J., & Paller, K. A. (2015). Functional differences between statistical learning with and without explicit training. Learning & Memory, 22(11), 544–556. https://doi.org/10.1101/lm.037986.114 148 Bauer, P. J., & Leventon, J. S. (2013). Memory for One-Time Experiences in the Second Year of Life: Implications for the Status of Episodic Memory. Infancy, 18(5), 755–781. https://doi.org/10.1111/infa.12005 Bays, B. C., Turk-Browne, N. B., & Seitz, A. R. (2015). Dissociable behavioural outcomes of visual statistical learning. Visual Cognition, 23(9–10), 1072–1097. https://doi.org/10.1080/13506285.2016.1139647 Bein, O., Trzewik, M., & Maril, A. (2019). The role of prior knowledge in incremental associative learning: An empirical and computational approach. Journal of Memory and Language, 107, 1–24. https://doi.org/10.1016/j.jml.2019.03.006 Bellana, B., Mansour, R., Ladyka-Wojcik, N., Grady, C., & Moscovitch, M. (2019). Prior knowledge facilitates recollection and the formation of durable memories. PsyArXiv. https://doi.org/10.31234/osf.io/wqa5y Bertels, J., Boursain, E., Destrebecqz, A., & Gaillard, V. (2015). Visual statistical learning in children and young adults: How implicit? Frontiers in Psychology, 5. https://doi.org/10.3389/fpsyg.2014.01541 Bertels, J., Destrebecqz, A., & Franco, A. (2015). Interacting Effects of Instructions and Presentation Rate on Visual Statistical Learning. Frontiers in Psychology, 6. https://doi.org/10.3389/fpsyg.2015.01806 Bertels, J., Franco, A., & Destrebecqz, A. (2012). How implicit is visual statistical learning? Journal of Experimental Psychology: Learning, Memory, and Cognition, 38(5), 1425–1431. https://doi.org/10.1037/a0027210 Bird, C. M., Davies, R. A., Ward, J., & Burgess, N. (2011). Effects of pre-experimental knowledge on recognition memory. Learning & Memory, 18(1), 11–14. https://doi.org/10.1101/lm.1952111 149 Bloxom, B. (1984). Estimating response time hazard functions: An exposition and extension. Journal of Mathematical Psychology, 28(4), 401–420. https://doi.org/10.1016/0022-2496(84)90008-7 Bogaerts, L., Siegelman, N., & Frost, R. (2016). Splitting the variance of statistical learning performance: A parametric investigation of exposure duration and transitional probabilities. Psychonomic Bulletin & Review, 23(4), 1250–1256. https://doi.org/10.3758/s13423-015-0996-z Bower, G. H., & Winzenz, D. (1970). Comparison of associative learning strategies. Psychonomic Science, 20(2), 119–120. https://doi.org/10.3758/BF03335632 Brady, T. F., Konkle, T., & Alvarez, G. A. (2009). Compression in visual working memory: Using statistical regularities to form more efficient memory representations. Journal of Experimental Psychology: General, 138(4), 487–502. Brady, T. F., & Oliva, A. (2008). Statistical Learning Using Real-World Scenes: Extracting Categorical Regularities Without Conscious Intent. Psychological Science, 19(7), 678–685. https://doi.org/10.1111/j.1467-9280.2008.02142.x Braine, M. D. S. (1963). On learning the grammatical order of words. Psychological Review, 70(4), 323–348. https://doi.org/10.1037/h0047696 Brod, G., Lindenberger, U., Werkle-Bergner, M., & Shing, Y. L. (2015). Differences in the neural signature of remembering schema-congruent and schema-incongruent events. NeuroImage, 117, 358–366. https://doi.org/10.1016/j.neuroimage.2015.05.086 Campbell, K. L., Zimerman, S., Healey, M. K., Lee, M. M. S., & Hasher, L. (2012). Age differences in visual statistical learning. Psychology and Aging, 27(3), 650–656. https://doi.org/10.1037/a0026780 Castello, M. V. di O., Halchenko, Y. O., Guntupalli, J. S., Gors, J. D., & Gobbini, M. I. (2017). The neural representation of personally familiar and unfamiliar faces in the distributed system for face perception. Scientific Reports, 7(1), 1–14. https://doi.org/10.1038/s41598-017-12559-1 150 Chait, M. (2020). How the brain discovers structure in sound sequences. Acoustical Science and Technology, 41(1), 48–53. https://doi.org/10.1250/ast.41.48 Chang, G. Y., & Knowlton, B. J. (2004). Visual Feature Learning in Artificial Grammar Classification. Journal of Experimental Psychology: Learning, Memory, and Cognition, 30(3), 714–722. https://doi.org/10.1037/0278-7393.30.3.714 Chomsky, N. (1976). On the Biological Basis of Language Capacities. In R. W. Rieber (Ed.), The Neuropsychology of Language: Essays in Honor of Eric Lenneberg (pp. 1–24). Springer US. https://doi.org/10.1007/978-1-4684-2292-4_1 Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204. https://doi.org/10.1017/S0140525X12000477 Cleeremans, A. (2006). Conscious and Unconscious Cognition: A Graded, Dynamic Perspective. International Journal of Psychology. Cleeremans, A., Destrebecqz, A., & Boyer, M. (1998). Implicit learning: News from the front. Trends in Cognitive Sciences, 2(10), 406–416. https://doi.org/10.1016/S1364-6613(98)01232-7 Conway, C. M., & Christiansen, M. H. (2006). Statistical Learning Within and Between Modalities: Pitting Abstract Against Stimulus-Specific Representations. Psychological Science, 17(10), 905–912. https://doi.org/10.1111/j.1467-9280.2006.01801.x Conway, Christopher M. (2020). How does the brain learn environmental structure? Ten core principles for understanding the neurocognitive mechanisms of statistical learning. Neuroscience & Biobehavioral Reviews, 112, 279–299. https://doi.org/10.1016/j.neubiorev.2020.01.032 Conway, Christopher M., & Christiansen, M. H. (2005). Modality-Constrained Statistical Learning of Tactile, Visual, and Auditory Sequences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(1), 24–39. https://doi.org/10.1037/0278-7393.31.1.24 151 Cooper, E., Greve, A., & Henson, R. N. (2019). Little evidence for Fast Mapping (FM) in adults: A review and discussion. Cognitive Neuroscience, 10(4), 196–209. https://doi.org/10.1080/17588928.2018.1542376 Coutanche, M. N., & Thompson-Schill, S. L. (2014). Fast mapping rapidly integrates information into existing memory networks. Journal of Experimental Psychology: General, 143(6), 2296–2303. https://doi.org/10.1037/xge0000020 Covington, N. V., Brown-Schmidt, S., & Duff, M. C. (2018). The Necessity of the Hippocampus for Statistical Learning. Journal of Cognitive Neuroscience, 30(5), 680–697. https://doi.org/10.1162/jocn_a_01228 Craik, F. I. M. (2002). Levels of processing: Past, present... and future? Memory, 10(5–6), 305–318. https://doi.org/10.1080/09658210244000135 Craik, F. I. M. (2020). Remembering: An Activity of Mind and Brain. Annual Review of Psychology, 71(1), 1–24. https://doi.org/10.1146/annurev-psych-010419-051027 Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 11(6), 671–684. https://doi.org/10.1016/S0022-5371(72)80001-X Craik, F. I. M., & Tulving, E. (1975). Depth of processing and the retention of words in episodic memory. Journal of Experimental Psychology: General, 104(3), 268–294. https://doi.org/10.1037/0096-3445.104.3.268 Cuddy, L., & Jacoby, L. (1982). When forgetting helps memory: An analysis of repetition effects. Journal of Verbal Learning and Verbal Behavior, 21(4), 451–467. https://doi.org/10.1016/S0022-5371(82)90727-7 Curran, T. (1997). Higher-order associative learning in amnesia: Evidence from the serial reaction time task. Journal of Cognitive Neuroscience, 9(4), 522–533. 152 Daltrozzo, J., & Conway, C. M. (2014). Neurocognitive mechanisms of statistical-sequential learning: What do event-related potentials tell us? Frontiers in Human Neuroscience, 8. https://doi.org/10.3389/fnhum.2014.00437 Davachi, L. (2006). Item, context and relational episodic encoding in humans. Current Opinion in Neurobiology, 16(6), 693–700. https://doi.org/10.1016/j.conb.2006.10.012 de Lange, F. P., Heilbron, M., & Kok, P. (2018). How Do Expectations Shape Perception? Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2018.06.002 de Lange, F. P., Jensen, O., & Dehaene, S. (2010). Accumulation of evidence during sequential decision making: The importance of top-down factors. J Neurosci, 30, 731–738. https://doi.org/10.1523/JNEUROSCI.4080-09.2010 de Leeuw, J. R. (2016). Dynamic constraints in statistical learning [Ph.D., Indiana University]. https://search.proquest.com/docview/1825682461/abstract/8223338BC3654729PQ/1 Degonda, N., Mondadori, C. R. A., Bosshardt, S., Schmidt, C. F., Boesiger, P., Nitsch, R. M., Hock, C., & Henke, K. (2005). Implicit Associative Learning Engages the Hippocampus and Interacts with Explicit Associative Learning. Neuron, 46(3), 505–520. https://doi.org/10.1016/j.neuron.2005.02.030 DiBartolo, M. M. (2019). “CAN’T LET IT GO”: THE PERSISTENCE OF CUE-DRIVEN REWARD-SEEKING DESPITE GOAL-RELATED CONSEQUENCES. [Thesis, Johns Hopkins University]. https://dspace-prod.mse.jhu.edu/handle/1774.2/62268 Eagle, M., & Leiter, E. (1964). Recall and recognition in intentional and incidental learning. Journal of Experimental Psychology, 68(1), 58–63. https://doi.org/10.1037/h0044655 Ebbinghaus, H. (1913). Retention as a function of the number of repetitions. In H. A. Ruger & C. E. Bussenius (Eds.), Memory: A contribution to experimental psychology (pp. 52–61). Teachers College Press. https://doi.org/10.1037/10011-006 153 Ellis, C. T., Skalaban, L. J., Yates, T. S., Bejjanki, V. R., Córdova, N. I., & Turk-Browne, N. B. (2020). Evidence of hippocampal learning in human infants. BioRxiv, 2020.10.07.329862. https://doi.org/10.1101/2020.10.07.329862 Ellis, N. C., & Sinclair, Susan G. (1996). Working Memory in the Acquisition of Vocabulary and Syntax: Putting Language in Good Order. The Quarterly Journal of Experimental Psychology Section A, 49(1), 234–250. https://doi.org/10.1080/713755604 Ellis, R. (2009). Implicit and explicit learning, knowledge and instruction. In Implicit and explicit knowledge in second language learning, testing and teaching. (pp. 3–25). Emberson, L. L., & Rubinstein, D. Y. (2016). Statistical learning is constrained to less abstract patterns in complex sensory input (but not the least). Cognition, 153, 63–78. https://doi.org/10.1016/j.cognition.2016.04.010 Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406. https://doi.org/10.1037/0033-295X.100.3.363 Ezzyat, Y., & Davachi, L. (2014). Similarity Breeds Proximity: Pattern Similarity within and across Contexts Is Related to Later Mnemonic Judgments of Temporal Proximity. Neuron, 81(5), 1179–1189. https://doi.org/10.1016/j.neuron.2014.01.042 Fernandes, T., Kolinsky, R., & Ventura, P. (2010). The impact of attention load on the use of statistical information and coarticulation as speech segmentation cues. Attention, Perception, & Psychophysics, 72(6), 1522–1532. https://doi.org/10.3758/APP.72.6.1522 Finn, A. S., & Hudson Kam, C. L. (2008). The curse of knowledge: First language knowledge impairs adult learners’ use of novel statistics for word segmentation. Cognition, 108(2), 477–499. https://doi.org/10.1016/j.cognition.2008.04.002 154 Finn, A. S., Kharitonova, M., Holtby, N., & Sheridan, M. A. (2018). Prefrontal and Hippocampal Structure Predict Statistical Learning Ability in Early Childhood. Journal of Cognitive Neuroscience, 31(1), 126–137. https://doi.org/10.1162/jocn_a_01342 Finn, A. S., Lee, T., Kraus, A., & Hudson Kam, C. L. (2014). When It Hurts (and Helps) to Try: The Role of Effort in Language Learning. PLoS ONE, 9(7), e101806. https://doi.org/10.1371/journal.pone.0101806 Fiser, J., & Aslin, R. N. (2001). Unsupervised Statistical Learning of Higher-Order Spatial Structures from Visual Scenes. Psychological Science, 12(6), 499–504. https://doi.org/10.1111/1467-9280.00392 Fiser, J., & Aslin, R. N. (2002). Statistical learning of higher-order temporal structure from visual shape sequences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 28(3), 458–467. https://doi.org/10.1037/0278-7393.28.3.458 Fiser, J., & Aslin, R. N. (2005). Encoding Multielement Scenes: Statistical Learning of Visual Feature Hierarchies. Journal of Experimental Psychology: General, 134(4), 521–537. https://doi.org/10.1037/0096-3445.134.4.521 Fiser, J., Scholl, B. J., & Aslin, R. N. (2007). Perceived object trajectories during occlusion constrain visual statistical learning. Psychonomic Bulletin & Review, 14(1), 173–178. https://doi.org/10.3758/BF03194046 Fletcher, P. C., Zafiris, O., Frith, C. D., Honey, R. A. E., Corlett, P. R., Zilles, K., & Fink, G. R. (2005). On the Benefits of not Trying: Brain Activity and Connectivity Reflecting the Interactions of Explicit and Implicit Sequence Learning. Cerebral Cortex, 15(7), 1002–1015. https://doi.org/10.1093/cercor/bhh201 Franco, A., Eberlen, J., Destrebecqz, A., Cleeremans, A., & Bertels, J. (2015). Rapid Serial Auditory Presentation. Experimental Psychology, 62(5), 346–351. https://doi.org/10.1027/1618-3169/a000295 155 Frost, R., Armstrong, B. C., & Christiansen, M. H. (2019). Statistical learning research: A critical review and possible new directions. Psychological Bulletin, No Pagination Specified-No Pagination Specified. https://doi.org/10.1037/bul0000210 Frost, R., Armstrong, B. C., Siegelman, N., & Christiansen, M. H. (2015). Domain generality versus modality specificity: The paradox of statistical learning. Trends in Cognitive Sciences, 19(3), 117–125. https://doi.org/10.1016/j.tics.2014.12.010 Fuster, J. M. (1997). Network memory. Trends in Neurosciences, 20(10), 451–459. https://doi.org/10.1016/S0166-2236(97)01128-4 Gathercole, S. E., Willis, C., Emslie, H., & Baddeley, A. D. (1991). The influences of number of syllables and wordlikeness on children’s repetition of nonwords. Applied Psycholinguistics, 12(3), 349–367. https://doi.org/10.1017/S0142716400009267 Gavornik, J. P., & Bear, M. F. (2014). Learned spatiotemporal sequence recognition and prediction in primary visual cortex. Nature Neuroscience, 17(5), 732–737. https://doi.org/10.1038/nn.3683 Gebhart, A. L., Aslin, R. N., & Newport, E. L. (2009). Changing Structures in Midstream: Learning Along the Statistical Garden Path. Cognitive Science, 33(6), 1087–1116. https://doi.org/10.1111/j.1551-6709.2009.01041.x Gilboa, A., & Marlatte, H. (2017). Neurobiology of Schemas and Schema-Mediated Memory. Trends in Cognitive Sciences, 21(8), 618–631. https://doi.org/10.1016/j.tics.2017.04.013 Glisky, E. L., Schacter, D. L., & Tulving, E. (1986). Computer learning by memory-impaired patients: Acquisition and retention of complex knowledge. Neuropsychologia, 24(3), 313–328. https://doi.org/10.1016/0028-3932(86)90017-5 Gmeindl, L., Walsh, M., & Courtney, S. M. (2011). Binding serial order to representations in working memory: A spatial/verbal dissociation. Memory & Cognition, 39(1), 37–46. https://doi.org/10.3758/s13421-010-0012-9 156 Goldrick, M., Folk, J. R., & Rapp, B. (2010). Mrs. Malaprop’s neighborhood: Using word errors to reveal neighborhood structure. Journal of Memory and Language, 62(2), 113–134. https://doi.org/10.1016/j.jml.2009.11.008 Gómez, R. L. (2002). Variability and Detection of Invariant Structure. Psychological Science, 13(5), 431–436. https://doi.org/10.1111/1467-9280.00476 Grill-Spector, K., & Malach, R. (2004). The human visual cortex. Annu Rev Neurosci, 27, 649–677. https://doi.org/10.1146/annurev.neuro.27.070203.144220 Grill-Spector, Kalanit, Henson, R., & Martin, A. (2006). Repetition and the brain: Neural models of stimulus-specific effects. Trends in Cognitive Sciences, 10(1), 14–23. https://doi.org/10.1016/j.tics.2005.11.006 Guillaume, C., Guillery-Girard, B., Chaby, L., Lebreton, K., Hugueville, L., Eustache, F., & Fiori, N. (2009). The time course of repetition effects for familiar faces and objects: An ERP study. Brain Research, 1248, 149–161. https://doi.org/10.1016/j.brainres.2008.10.069 Gureckis, T., & Love, B. (2007). Behaviorism reborn? Statistical learning as simple conditioning. Proceedings of the Annual Meeting of the Cognitive Science Society, 29(29), 335–340. Gureckis, T. M., Martin, J., McDonnell, J., Rich, A. S., Markant, D., Coenen, A., Halpern, D., Hamrick, J. B., & Chan, P. (2016). psiTurk: An open-source framework for conducting replicable behavioral experiments online. Behavior Research Methods, 48(3), 829–842. https://doi.org/10.3758/s13428-015-0642-8 Hasher, L., & Zacks, R. T. (1979). Automatic and effortful processes in memory. Journal of Experimental Psychology: General, 108(3), 356–388. https://doi.org/10.1037/0096-3445.108.3.356 Hasson, U., Yang, E., Vallines, I., Heeger, D. J., & Rubin, N. (2008). A Hierarchy of Temporal Receptive Windows in Human Cortex. Journal of Neuroscience, 28(10), 2539–2550. https://doi.org/10.1523/JNEUROSCI.5487-07.2008 157 Hasson, Uri, Chen, J., & Honey, C. J. (2015). Hierarchical process memory: Memory as an integral component of information processing. Trends in Cognitive Sciences, 19(6), 304–313. https://doi.org/10.1016/j.tics.2015.04.006 Healey, M. K. (2018). Temporal contiguity in incidentally encoded memories. Journal of Memory and Language, 102, 28–40. https://doi.org/10.1016/j.jml.2018.04.003 Hebb, D. O. (1949). The organization of behavior: A neuropsychological theory. J. Wiley; Chapman & Hall. Heilbron, M., & Chait, M. (2017). Great expectations: Is there evidence for predictive coding in auditory cortex? Neuroscience. https://doi.org/10.1016/j.neuroscience.2017.07.061 Henke, K. (2010). A model for memory systems based on processing modes rather than consciousness. Nature Reviews Neuroscience, 11(7), 523–532. https://doi.org/10.1038/nrn2850 Himberger, K. D., Chien, H.-Y., & Honey, C. J. (2018). Principles of Temporal Processing Across the Cortical Hierarchy. Neuroscience. https://doi.org/10.1016/j.neuroscience.2018.04.030 Himberger, K. D., Finn, A., & Honey, C. J. (2019). Reconsidering the Automaticity of Visual Statistical Learning [Preprint]. PsyArXiv. https://doi.org/10.31234/osf.io/r659w Hogan, K., & Pressley, M. (Eds.). (1997). Scaffolding student learning: Instructional approaches and issues (p. 198). Brookline Books. Honey, C. J., Thesen, T., Donner, T. H., Silbert, L. J., Carlson, C. E., Devinsky, O., Doyle, W. K., Rubin, N., Heeger, D. J., & Hasson, U. (2012). Slow Cortical Dynamics and the Accumulation of Information over Long Timescales. Neuron, 76(2), 423–434. https://doi.org/10.1016/j.neuron.2012.08.011 Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. The Journal of Physiology, 160(1), 106. Hunt, R. R., & Einstein, G. (1981). Relational and item-specific information in memory. Journal of Verbal Learning and Verbal Behavior, 20(5), 497–514. https://doi.org/10.1016/S0022-5371(81)90138-9 158 Hussain, M. M., & Mahmud, I. (2019). pyMannKendall: A python package for non parametric Mann Kendall family of trend tests. Journal of Open Source Software, 4(39), 1556. https://doi.org/10.21105/joss.01556 Jiang, Y., & Chun, M. M. (2001). Selective attention modulates implicit learning. The Quarterly Journal of Experimental Psychology Section A, 54(4), 1105–1124. https://doi.org/10.1080/713756001 Jiménez, L., & Méndez, C. (1999). Which attention is needed for implicit sequence learning? Journal of Experimental Psychology: Learning, Memory, and Cognition, 25(1), 236–259. https://doi.org/10.1037/0278-7393.25.1.236 Jones, E., Oliphant, T., Peterson, P., & others. (2001). SciPy: Open source scientific tools for Python. http://www.scipy.org Kahana, M. J. (1996). Associative retrieval processes in free recall. Memory & Cognition, 24(1), 103–109. https://doi.org/10.3758/BF03197276 Kanwisher, N. (2010). Functional specificity in the human brain: A window into the functional architecture of the mind. Proceedings of the National Academy of Sciences, 107(25), 11163–11170. https://doi.org/10.1073/pnas.1005062107 Karuza, E. A., Emberson, L. L., & Aslin, R. N. (2014). Combining fMRI and behavioral measures to examine the process of human learning. Neurobiology of Learning and Memory, 109, 193–206. https://doi.org/10.1016/j.nlm.2013.09.012 Karuza, E. A., Newport, E. L., Aslin, R. N., Starling, S. J., Tivarus, M. E., & Bavelier, D. (2013). The neural correlates of statistical learning in a word segmentation task: An fMRI study. Brain and Language, 127(1), 46–54. https://doi.org/10.1016/j.bandl.2012.11.007 Kerby, D. S. (2014). The Simple Difference Formula: An Approach to Teaching Nonparametric Correlation. Comprehensive Psychology, 3, 11.IT.3.1. https://doi.org/10.2466/11.IT.3.1 159 Kesteren, M. T. R. van, Krabbendam, L., & Meeter, M. (2018). Integrating educational knowledge: Reactivation of prior knowledge during educational learning enhances memory integration. Npj Science of Learning, 3(1), 11. https://doi.org/10.1038/s41539-018-0027-8 Kidd, C., Piantadosi, S. T., & Aslin, R. N. (2012). The Goldilocks Effect: Human Infants Allocate Attention to Visual Sequences That Are Neither Too Simple Nor Too Complex. PLoS ONE, 7(5), 1–8. https://doi.org/10.1371/journal.pone.0036399 Kim, R., Seitz, A., Feenstra, H., & Shams, L. (2009). Testing assumptions of statistical learning: Is it long-term and implicit? Neuroscience Letters, 461(2), 145–149. https://doi.org/10.1016/j.neulet.2009.06.030 Kirkham, N. Z., Slemmer, J. A., & Johnson, S. P. (2002). Visual statistical learning in infancy: Evidence for a domain general learning mechanism. Cognition, 83(2), B35–B42. https://doi.org/10.1016/S0010-0277(02)00004-5 Klatzky, R. L., & Forrest, F. H. (1984). Recognizing familiar and unfamiliar faces. Memory & Cognition, 12(1), 60–70. https://doi.org/10.3758/BF03196998 Kok, P., Mostert, P., & Lange, F. P. de. (2017). Prior expectations induce prestimulus sensory templates. Proceedings of the National Academy of Sciences, 201705652. https://doi.org/10.1073/pnas.1705652114 Kole, J. A., & Healy, A. F. (2007). Using prior knowledge to minimize interference when learning large amounts of information. Memory & Cognition, 35(1), 124–137. https://doi.org/10.3758/BF03195949 Körding, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature, 427(6971), 244–247. https://doi.org/10.1038/nature02169 160 Kraut, R., Olson, J., Banaji, M., Bruckman, A., Cohen, J., & Couper, M. (2004). Psychological Research Online: Report of Board of Scientific Affairs’ Advisory Group on the Conduct of Research on the Internet. American Psychologist, 59(2), 105–117. https://doi.org/10.1037/0003-066X.59.2.105 Kremláček, J., Kreegipuu, K., Tales, A., Astikainen, P., Põldver, N., Näätänen, R., & Stefanics, G. (2016). Visual mismatch negativity (vMMN): A review and meta-analysis of studies in psychiatric and neurological disorders. Cortex, 80, 76–112. https://doi.org/10.1016/j.cortex.2016.03.017 Kutas, M., & Federmeier, K. D. (2011). Thirty Years and Counting: Finding Meaning in the N400 Component of the Event-Related Brain Potential (ERP). Annual Review of Psychology, 62(1), 621–647. https://doi.org/10.1146/annurev.psych.093008.131123 Le Pelley, M. E., Mitchell, C. J., Beesley, T., George, D. N., & Wills, A. J. (2016). Attention and associative learning in humans: An integrative review. Psychological Bulletin, 142(10), 1111–1140. https://doi.org/10.1037/bul0000064 Lerner, Y., Honey, C. J., Silbert, L. J., & Hasson, U. (2011). Topographic Mapping of a Hierarchy of Temporal Receptive Windows Using a Narrated Story. Journal of Neuroscience, 31(8), 2906–2915. https://doi.org/10.1523/JNEUROSCI.3684-10.2011 Liu, Z.-X., Grady, C., & Moscovitch, M. (2017). Effects of Prior-Knowledge on Brain Activation and Connectivity During Associative Memory Encoding. Cerebral Cortex (New York, N.Y.: 1991), 27(3), 1991–2009. https://doi.org/10.1093/cercor/bhw047 Luo, Y., & Zhao, J. (2018). Statistical Learning Creates Novel Object Associations via Transitive Relations. Psychological Science, 29(8), 1207–1220. https://doi.org/10.1177/0956797618762400 Marcus, G. F., Vijayan, S., Rao, S. B., & Vishton, P. M. (1999). Rule Learning by Seven-Month-Old Infants. Science, 283(5398), 77–80. https://doi.org/10.1126/science.283.5398.77 161 May, P. J. C., & Tiitinen, H. (2010). Mismatch negativity (MMN), the deviance-elicited auditory deflection, explained. Psychophysiology, 47(1), 66–122. https://doi.org/10.1111/j.1469-8986.2009.00856.x Mayes, A., Montaldi, D., & Migo, E. (2007). Associative memory and the medial temporal lobes. Trends in Cognitive Sciences, 11(3), 126–135. https://doi.org/10.1016/j.tics.2006.12.003 McClelland, J. L., McNaughton, B., & O’Reilly, R. C. (1995). Why There Are Complementary Learning Systems in the Hippocampus and Neocortex: Insights from the Successes and Failures of Connectionist Models of Learning and Memory. Psychological Review, 102(3,419-457). McNealy, K., Mazziotta, J. C., & Dapretto, M. (2006). Cracking the Language Code: Neural Mechanisms Underlying Speech Parsing. Journal of Neuroscience, 26(29), 7629–7639. https://doi.org/10.1523/JNEUROSCI.5501-05.2006 Megreya, A. M., & Burton, A. M. (2006). Unfamiliar faces are not faces: Evidence from a matching task. Memory & Cognition, 34(4), 865–876. https://doi.org/10.3758/BF03193433 Mesulam, M. M. (1998). From sensation to cognition. Brain, 121 ( Pt 6), 1013–1052. Mickes, L., Wais, P. E., & Wixted, J. T. (2009). Recollection Is a Continuous Process: Implications for Dual-Process Theories of Recognition Memory. Psychological Science, 20(4), 509–515. https://doi.org/10.1111/j.1467-9280.2009.02324.x Murphy, K. R., Myors, B., & Wolach, A. (2014). Statistical Power Analysis: A Simple and General Model for Traditional and Modern Hypothesis Tests, Fourth Edition. Routledge. https://doi.org/10.4324/9781315773155 Murphy, R. A., Mondragón, E., & Murphy, V. A. (2008). Rule Learning by Rats. Science, 319(5871), 1849–1851. https://doi.org/10.1126/science.1151564 162 Musen, G., & Treisman, A. (1990). Implicit and explicit memory for visual patterns. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(1), 127–137. https://doi.org/10.1037/0278-7393.16.1.127 Musz, E., Weber, M. J., & Thompson-Schill, S. L. (2015). Visual statistical learning is not reliably modulated by selective attention to isolated events. Attention, Perception, & Psychophysics, 77(1), 78–96. https://doi.org/10.3758/s13414-014-0757-5 Näätänen, R., Gaillard, A. W. K., & Mäntysalo, S. (1978). Early selective-attention effect on evoked potential reinterpreted. Acta Psychologica, 42(4), 313–329. https://doi.org/10.1016/0001-6918(78)90006-9 Näätänen, R., Paavilainen, P., Rinne, T., & Alho, K. (2007). The mismatch negativity (MMN) in basic research of central auditory processing: A review. Clinical Neurophysiology, 118(12), 2544–2590. https://doi.org/10.1016/j.clinph.2007.04.026 Näätänen, Risto, Lehtokoski, A., Lennes, M., Cheour, M., Huotilainen, M., Iivonen, A., Vainio, M., Alku, P., Ilmoniemi, R. J., Luuk, A., Allik, J., Sinkkonen, J., & Alho, K. (1997). Language-specific phoneme representations revealed by electric and magnetic brain responses. Nature, 385(6615), 432–434. https://doi.org/10.1038/385432a0 Nadel, L., & Zola-Morgan, S. (1984). Infantile Amnesia. In M. Moscovitch (Ed.), Infant Memory: Its Relation to Normal and Pathological Memory in Humans and Other Animals (pp. 145–172). Springer US. https://doi.org/10.1007/978-1-4615-9364-5_7 Ngiam, W. X. Q., Brissenden, J. A., & Awh, E. (2019). “Memory compression” effects in visual working memory are contingent on explicit long-term memory. Journal of Experimental Psychology: General, 148(8), 1373–1385. https://doi.org/10.1037/xge0000649 Nissen, M. J., & Bullemer, P. (1987). Attentional requirements of learning: Evidence from performance measures. Cognitive Psychology, 19(1), 1–32. https://doi.org/10.1016/0010-0285(87)90002-8 163 O’Keefe, J., & Nadel, L. (1978). The hippocampus as a cognitive map. Clarendon Press ; Oxford University Press. Olson, I. R., & Chun, M. M. (2001). Temporal contextual cuing of visual attention. Journal of Experimental Psychology: Learning, Memory, and Cognition, 27(5), 1299–1313. https://doi.org/10.1037/0278-7393.27.5.1299 Otsuka, S., Koch, C., & Saiki, J. (2016). Visual statistical learning produces implicit and explicit knowledge about temporal order information and scene chunks: Evidence from direct and indirect measures. Visual Cognition, 24(2), 155–172. https://doi.org/10.1080/13506285.2016.1211209 Otsuka, S., Nishiyama, M., Nakahara, F., & Kawaguchi, J. (2013). Visual statistical learning based on the perceptual and semantic information of objects. Journal of Experimental Psychology: Learning, Memory, and Cognition, 39(1), 196–207. https://doi.org/10.1037/a0028645 Otsuka, S., & Saiki, J. (2017). Neural correlates of implicit knowledge about statistical regularities. Experimental Brain Research, 235. https://doi.org/10.1007/s00221-017-5083-4 Pacton, S., & Perruchet, P. (2008). An attention-based associative account of adjacent and nonadjacent dependency learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 34(1), 80–96. https://doi.org/10.1037/0278-7393.34.1.80 Park, S. H., Rogers, L. L., & Vickery, T. J. (2018). The roles of order, distance, and interstitial items in temporal visual statistical learning. Attention, Perception, & Psychophysics, 80(6), 1409–1419. https://doi.org/10.3758/s13414-018-1556-1 Peretz, I., Saffran, J. R., Schön, D., & Gosselin, N. (2012). Statistical learning of speech, not music, in congenital amusia: Peretz et al. Annals of the New York Academy of Sciences, 1252(1), 361–366. https://doi.org/10.1111/j.1749-6632.2011.06429.x Perfors, A., & Kidd, E. (2020). What drives individual differences in statistical learning? The role of perceptual fluency and familiarity. https://doi.org/10.31234/osf.io/aw3pq 164 Perrachione, T. K., Ghosh, S. S., Ostrovskaya, I., Gabrieli, J. D. E., & Kovelman, I. (2017). Phonological Working Memory for Words and Nonwords in Cerebral Cortex. Journal of Speech, Language, and Hearing Research, 60(7), 1959–1979. https://doi.org/10.1044/2017_JSLHR-L-15-0446 Perruchet, P., & Pacton, S. (2006). Implicit learning and statistical learning: One phenomenon, two approaches. Trends Cogn Sci, 10, 233–238. https://doi.org/10.1016/j.tics.2006.03.006 Perruchet, Pierre, Cleeremans, A., & Destrebecqz, A. (2006). Dissociating the effects of automatic activation and explicit expectancy on reaction times in a simple associative learning task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32(5), 955–965. https://doi.org/10.1037/0278-7393.32.5.955 Perruchet, Pierre, & Pacton, S. (2006). Implicit learning and statistical learning: One phenomenon, two approaches. Trends in Cognitive Sciences, 10(5), 233–238. https://doi.org/10.1016/j.tics.2006.03.006 Plante, E., & Gómez, R. L. (2018). Learning Without Trying: The Clinical Relevance of Statistical Learning. Language, Speech, and Hearing Services in Schools, 49(3S), 710–722. https://doi.org/10.1044/2018_LSHSS-STLT1-17-0131 Prince, S. E., Daselaar, S. M., & Cabeza, R. (2005). Neural Correlates of Relational Memory: Successful Encoding and Retrieval of Semantic and Perceptual Associations. Journal of Neuroscience, 25(5), 1203–1210. https://doi.org/10.1523/JNEUROSCI.2540-04.2005 Ramsaran, A. I., Schlichting, M. L., & Frankland, P. W. (2019). The ontogeny of memory persistence and specificity. Developmental Cognitive Neuroscience, 36, 100591. https://doi.org/10.1016/j.dcn.2018.09.002 Rauschecker, J. P., & Scott, S. K. (2009). Maps and streams in the auditory cortex: Nonhuman primates illuminate human speech processing. Nature Neuroscience, 12(6), 718–724. https://doi.org/10.1038/nn.2331 165 Rawal, A., & Tseng, P. (2021). The effect of visual statistical learning in RSVP: Implicit learning or stream location artifact? Journal of Experimental Psychology: Learning, Memory, and Cognition, No Pagination Specified-No Pagination Specified. https://doi.org/10.1037/xlm0000988 Raz, G., & Saxe, R. (2020). Learning in Infancy Is Active, Endogenously Motivated, and Depends on the Prefrontal Cortices. Annual Review of Developmental Psychology, 2(1), null. https://doi.org/10.1146/annurev-devpsych-121318-084841 Reber, A. S. (1967). Implicit Learning of Artifical Grammars. Journal of Verbal Learning and Verbal Behavior, 6, 855–863. Reber, P. J. (2013). The neural basis of implicit learning and memory: A review of neuropsychological and neuroimaging research. Neuropsychologia, 51(10), 2026–2042. https://doi.org/10.1016/j.neuropsychologia.2013.06.019 Reder, L. M., Park, H., & Kieffaber, P. (2009). Memory systems do not divide on consciousness: Reinterpreting memory in terms of activation and binding. Psychological Bulletin, 135(1), 23–49. https://doi.org/10.1037/a0013974 Reder, L. M., Victoria, L. W., Manelis, A., Oates, J. M., Dutcher, J. M., Bates, J. T., Cook, S., Aizenstein, H. J., Quinlan, J., & Gyulai, F. (2013). Why It’s Easier to Remember Seeing a Face We Already Know Than One We Don’t: Preexisting Memory Representations Facilitate Memory Formation. Psychological Science, 24(3), 363–372. https://doi.org/10.1177/0956797612457396 Richter, D., & de Lange, F. P. (2019). Statistical learning attenuates visual activity only for attended stimuli. ELife, 8, e47869. https://doi.org/10.7554/eLife.47869 Saffran, J. R. (2003). Statistical Language Learning: Mechanisms and Constraints. Current Directions in Psychological Science, 12(4), 110–114. https://doi.org/10.1111/1467-8721.01243 Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science (New York, N.Y.), 274(5294), 1926–1928. 166 Saffran, J. R., Johnson, E. K., Aslin, R. N., & Newport, E. L. (1999a). Statistical learning of tone sequences by human infants and adults. Cognition, 70(1), 27–52. https://doi.org/10.1016/S0010-0277(98)00075-4 Saffran, J. R., Johnson, E. K., Aslin, R. N., & Newport, E. L. (1999b). Statistical learning of tone sequences by human infants and adults. Cognition, 70(1), 27–52. Saffran, J. R., & Kirkham, N. Z. (2018). Infant Statistical Learning. Annual Review of Psychology, 69(1), 181–203. https://doi.org/10.1146/annurev-psych-122216-011805 Saffran, J. R., Newport, E. L., & Aslin, R. N. (1996). Word Segmentation: The Role of Distributional Cues. Journal of Memory and Language, 35(4), 606–621. https://doi.org/10.1006/jmla.1996.0032 Saffran, J. R., Newport, E. L., Aslin, R. N., Tunick, R. A., & Barrueco, S. (1997). Incidental Language Learning: Listening (and Learning) Out of the Corner of Your Ear. Psychological Science, 8(2), 101–105. https://doi.org/10.1111/j.1467-9280.1997.tb00690.x Schacter, D. L. (1987). Implicit Memory: History and Current Status. Journal of Experimental Psychology, 13(3), 501–518. Schacter, D. L. (2019). Implicit Memory, Constructive Memory, and Imagining the Future: A Career Perspective. Perspectives on Psychological Science, 14(2), 256–272. https://doi.org/10.1177/1745691618803640 Schacter, D. L., & Graf, P. (1989). Modality specificity of implicit memory for new associations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15(1), 3–12. https://doi.org/10.1037/0278-7393.15.1.3 Schacter, D. L., & McGlynn, S. M. (1989). Implicit Memory: Effects of Elaboration Depend on Unitization. The American Journal of Psychology, 102(2), 151–181. https://doi.org/10.2307/1422950 167 Schapiro, A. C., Gregory, E., Landau, B., McCloskey, M., & Turk-Browne, N. B. (2014). The Necessity of the Medial Temporal Lobe for Statistical Learning. Journal of Cognitive Neuroscience, 26(8), 1736–1747. https://doi.org/10.1162/jocn_a_00578 Schapiro, A. C., Kustner, L. V., & Turk-Browne, N. B. (2012). Shaping of Object Representations in the Human Medial Temporal Lobe Based on Temporal Regularities. Current Biology, 22(17), 1622–1627. https://doi.org/10.1016/j.cub.2012.06.056 Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne, N. B., & Botvinick, M. M. (2013). Neural representations of events arise from temporal community structure. Nature Neuroscience, 16(4), 486–492. https://doi.org/10.1038/nn.3331 Schapiro, A. C., Turk-Browne, N. B., Norman, K. A., & Botvinick, M. M. (2016). Statistical learning of temporal community structure in the hippocampus. Hippocampus, 26(1), 3–8. https://doi.org/10.1002/hipo.22523 Schlichting, M. L., Guarino, K. F., Schapiro, A. C., Turk-Browne, N. B., & Preston, A. R. (2017). Hippocampal Structure Predicts Statistical Learning and Associative Inference Abilities during Development. Journal of Cognitive Neuroscience, 29(1), 37–51. https://doi.org/10.1162/jocn_a_01028 Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. Schwartz, L., & Yovel, G. (2016). The roles of perceptual and conceptual information in face recognition. Journal of Experimental Psychology: General, 145(11), 1493–1511. https://doi.org/10.1037/xge0000220 Schwartz, L., & Yovel, G. (2019). Learning faces as concepts rather than percepts improves face recognition. Journal of Experimental Psychology. Learning, Memory, and Cognition, 45(10), 1733–1747. https://doi.org/10.1037/xlm0000673 168 Seabold, S., & Perktold, J. (2010). Statsmodels: Econometric and Statistical Modeling with Python. 92–96. https://doi.org/10.25080/Majora-92bf1922-011 Shohamy, D., & Turk-Browne, N. B. (2013). Mechanisms for widespread hippocampal involvement in cognition. Journal of Experimental Psychology: General, 142(4), 1159–1170. https://doi.org/10.1037/a0034461 Shufaniya, A., & Arnon, I. (2018). Statistical Learning Is Not Age-Invariant During Childhood: Performance Improves With Age Across Modality. Cognitive Science, 42(8), 3100–3115. https://doi.org/10.1111/cogs.12692 Siegelman, N., Bogaerts, L., Elazar, A., Arciuli, J., & Frost, R. (2018). Linguistic entrenchment: Prior knowledge impacts statistical learning performance. Cognition, 177, 198–213. https://doi.org/10.1016/j.cognition.2018.04.011 Siegelman, N., Bogaerts, L., & Frost, R. (2017). Measuring individual differences in statistical learning: Current pitfalls and possible solutions. Behavior Research Methods, 49(2), 418–432. https://doi.org/10.3758/s13428-016-0719-z Siegelman, N., Bogaerts, L., Kronenfeld, O., & Frost, R. (2018). Redefining “Learning” in Statistical Learning: What Does an Online Measure Reveal About the Assimilation of Visual Regularities? Cognitive Science, 42(S3), 692–727. https://doi.org/10.1111/cogs.12556 Siegelman, N., & Frost, R. (2015). Statistical learning as an individual ability: Theoretical perspectives and empirical evidence. Journal of Memory and Language, 81, 105–120. https://doi.org/10.1016/j.jml.2015.02.001 Song, S., Howard, J. H., & Howard, D. V. (2007). Implicit probabilistic sequence learning is independent of explicit awareness. Learning & Memory, 14(3), 167–176. https://doi.org/10.1101/lm.437407 Song, S., Howard Jr., J. H., & Howard, D. V. (2008). Perceptual sequence learning in a serial reaction time task. Experimental Brain Research, 189(2), 145–158. https://doi.org/10.1007/s00221-008-1411-z 169 Spelke, E. S., & Newport, E. L. (1998). Nativism, empiricism, and the development of knowledge. In Handbook of child psychology: Theoretical models of human development, Volume 1, 5th ed (pp. 275–340). John Wiley & Sons Inc. Squire, L. R., & Zola, S. M. (1996). Structure and function of declarative and nondeclarative memory systems. Proceedings of the National Academy of Sciences, 93(24), 13515–13522. https://doi.org/10.1073/pnas.93.24.13515 Stall, S., Yarmey, L., Cutcher-Gershenfeld, J., Hanson, B., Lehnert, K., Nosek, B., Parsons, M., Robinson, E., & Wyborn, L. (2019). Make scientific data FAIR. Nature, 570(7759), 27–29. https://doi.org/10.1038/d41586-019-01720-7 Stark, C. E. L., & Okado, Y. (2003). Making Memories without Trying: Medial Temporal Lobe Activity Associated with Incidental Memory Formation during Recognition. Journal of Neuroscience, 23(17), 6748–6753. https://doi.org/10.1523/JNEUROSCI.23-17-06748.2003 Stevens, D. J., Arciuli, J., & Anderson, D. I. (2014). Concurrent Movement Impairs Incidental But Not Intentional Statistical Learning. Cognitive Science, 39(5), 1081–1098. https://doi.org/10.1111/cogs.12180 Tacikowski, P., Jednoróg, K., Marchewka, A., & Nowicka, A. (2011). How multiple repetitions influence the processing of self-, famous and unknown names and faces: An ERP study. International Journal of Psychophysiology, 79(2), 219–230. https://doi.org/10.1016/j.ijpsycho.2010.10.010 Thiessen, E. D. (2010). Effects of Visual Information on Adults’ and Infants’ Auditory Statistical Learning. Cognitive Science, 34(6), 1093–1106. https://doi.org/10.1111/j.1551-6709.2010.01118.x Thiessen, E. D. (2017). What’s statistical about learning? Insights from modelling statistical learning as a set of memory processes. Phil. Trans. R. Soc. B, 372(1711), 20160056. https://doi.org/10.1098/rstb.2016.0056 170 Thorndike, E. L. (1898). Animal intelligence: An experimental study of the associative processes in animals. The Psychological Review: Monograph Supplements, 2(4), i–109. https://doi.org/10.1037/h0092987 Todorovic, A., Ede, F. van, Maris, E., & Lange, F. P. de. (2011). Prior Expectation Mediates Neural Adaptation to Repeated Sounds in the Auditory Cortex: An MEG Study. Journal of Neuroscience, 31(25), 9118–9123. https://doi.org/10.1523/JNEUROSCI.1425-11.2011 Tomczak, M., & Tomczak, E. (2014). The need to report effect size estimates revisited. An overview of some recommended measures of effect size. Trends in Sports Sciences, 1(21), 19–25. Toro, J. M., Sinnett, S., & Soto-Faraco, S. (2005). Speech segmentation by statistical learning depends on attention. Cognition, 97(2), B25–B34. https://doi.org/10.1016/j.cognition.2005.01.006 Turk-Browne, N. B. (2012). Statistical Learning and Its Consequences. In The Influence of Attention, Learning, and Motivation on Visual Search (pp. 117–146). Springer, New York, NY. https://doi.org/10.1007/978-1-4614-4794-8_6 Turk-Browne, N. B., Jungé, J. A., & Scholl, B. J. (2005). The Automaticity of Visual Statistical Learning. Journal of Experimental Psychology: General, 134(4), 552–564. https://doi.org/10.1037/0096-3445.134.4.552 Turk-Browne, N. B., & Scholl, B. J. (2009). Flexible visual statistical learning: Transfer across space and time. Journal of Experimental Psychology: Human Perception and Performance, 35(1), 195–202. https://doi.org/10.1037/0096-1523.35.1.195 Turk-Browne, N. B., Scholl, B. J., Chun, M. M., & Johnson, M. K. (2009). Neural evidence of statistical learning: Efficient detection of visual regularities without awareness. Journal of Cognitive Neuroscience, 21(10), 1934–1945. 171 Turk-Browne, N. B., Scholl, B. J., Johnson, M. K., & Chun, M. M. (2010). Implicit Perceptual Anticipation Triggered by Statistical Learning. Journal of Neuroscience, 30(33), 11177–11187. https://doi.org/10.1523/JNEUROSCI.0858-10.2010 Underwood, B. J. (1969). Some correlates of item repetition in free-recall learning. Journal of Verbal Learning and Verbal Behavior, 8(1), 83–94. https://doi.org/10.1016/S0022-5371(69)80015-0 Vakil, E., Kahan, S., Huberman, M., & Osimani, A. (2000). Motor and non-motor sequence learning in patients with basal ganglia lesions: The case of serial reaction time (SRT). Neuropsychologia, 38(1), 1–10. https://doi.org/10.1016/S0028-3932(99)00058-5 Vallet, R. (2018). Pingouin: Statistics in Python. The Journal of Open Source Software, 3(31), 1026. van Kesteren, M. T. R., Beul, S. F., Takashima, A., Henson, R. N., Ruiter, D. J., & Fernández, G. (2013). Differential roles for medial prefrontal and medial temporal cortices in schema-dependent encoding: From congruent to incongruent. Neuropsychologia, 51(12), 2352–2359. https://doi.org/10.1016/j.neuropsychologia.2013.05.027 van Kesteren, M. T. R., Ruiter, D. J., Fernández, G., & Henson, R. N. (2012). How schema and novelty augment memory formation. Trends in Neurosciences, 35(4), 211–219. https://doi.org/10.1016/j.tins.2012.02.001 Vergnieux, V., & Vogels, R. (2020). Statistical Learning Signals for Complex Visual Images in Macaque Early Visual Cortex. Frontiers in Neuroscience. https://doi.org/10.3389/fnins.2020.00789 Vickery, T., Kuhl, B., & Chun, M. (2012). Temporally specific visual working memory representations revealed by multivoxel pattern analysis. Journal of Vision, 12(9), 1272–1272. https://doi.org/10.1167/12.9.1272 Vickery, T., Park, S. H., Gupta, J., & Berryhill, M. (2018). Tasks determine what is learned in visual statistical learning. Psychonomic Bulletin & Review, 25(5), 1847–1854. https://doi.org/10.3758/s13423-017-1405-6 172 Vinken, K., Vogels, R., & Op de Beeck, H. (2017). Recent Visual Experience Shapes Visual Processing in Rats through Stimulus-Specific Adaptation and Response Enhancement. Current Biology, 27(6), 914–919. https://doi.org/10.1016/j.cub.2017.02.024 Wagner, L. (2001). Defending nativism in language acquisition. Trends Cogn Sci, 5, 283–284. Xu, J., Kemeny, S., Park, G., Frattali, C., & Braun, A. (2005). Language in context: Emergent features of word, sentence, and narrative comprehension. Neuroimage, 25, 1002–1015. https://doi.org/10.1016/j.neuroimage.2004.12.013 Yang, J., & Li, P. (2012). Brain Networks of Explicit and Implicit Learning. PLOS ONE, 7(8), e42993. https://doi.org/10.1371/journal.pone.0042993 Yates, T. S., Ellis, C. T., & Turk-Browne, N. B. (2021). The promise of awake behaving infant fMRI as a deep measure of cognition. Current Opinion in Behavioral Sciences, 40, 5–11. https://doi.org/10.1016/j.cobeha.2020.11.007 Yonelinas, A. (1999). The contribution of recollection and familiarity to recognition and source-memory judgments: A formal dual-process model and an analysis of receiver operating characterstics. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25(6), 1415–1434. Yonelinas, A. P., Kroll, N. E. A., Dobbins, I. G., & Soltani, M. (1999). Recognition memory for faces: When familiarity supports associative recognition judgments. Psychonomic Bulletin & Review, 6(4), 654–661. https://doi.org/10.3758/BF03212975 Yonelinas, Andrew P., & Jacoby, L. L. (2012). The process-dissociation approach two decades later: Convergence, boundary conditions, and new directions. Memory & Cognition, 40(5), 663–680. https://doi.org/10.3758/s13421-012-0205-5 Yott, J., & Poulin-Dubois, D. (2016). Are Infants’ Theory-of-Mind Abilities Well Integrated? Implicit Understanding of Intentions, Desires, and Beliefs. Journal of Cognition and Development, 17(5), 683–698. https://doi.org/10.1080/15248372.2015.1086771 173 Yurovsky, D., Yu, C., & Smith, L. B. (2012). Statistical speech segmentation and word learning in parallel: Scaffolding from child-directed speech. Front Psychol, 3, 374. https://doi.org/10.3389/fpsyg.2012.00374 Zhao, J., Al-Aidroos, N., & Turk-Browne, N. B. (2013). Attention is spontaneously biased toward regularities. Psychological Science, 0956797612460407. Zhou, Z., Singh, D., Tandoc, M., & Schapiro, A. C. (2020). Interleaving facilitates the rapid formation of distributed representations. Proceedings of the 42nd Annual Conference of the Cognitive Science Society. Ziori, E., & Dienes, Z. (2008). How does Prior Knowledge Affect Implicit and Explicit Concept Learning? Quarterly Journal of Experimental Psychology, 61(4), 601–624. https://doi.org/10.1080/17470210701255374 174 Supplementary Materials Supplemental Methods Participant Demographics Supplemental Figure 1: Age & gender statistics. (a) Distribution of age groups by condition. (b) Distribution of gender by condition. Data Collection Statistics 175 Supplemental Table 1: Data collection statistics across all conditions Attend Jiggle (0 Repetitions) Attend Jiggle (60 Repetitions) Attend Jiggle (96 Repetitions) 1-back Attend Stimuli Attend Relationships 2-back Attend Triplets Feedback Training Traditional TD Scrambled TD Non-famous Faces Famous Faces Experiment # 1 1 1 2 2 2 2 2 2 3 3 4 4 Cover Task Detect Image Motion Detect Image Motion Detect Image Motion Detect repeat images Attend Stimuli Attend Relates Detect repeat images Attend Triplet Structure Explicit Feedback Training Detect Image Motion Detect Image Motion Detect Repeats Detect Repeats Stimulus Type Fractals Fractals Fractals Fractals Fractals Fractals Fractals Fractals Fractals Fractals Fractals Novel Faces Famous Faces # Participants Analyzed 120 120 120 120 120 120 120 120 54 120 120 60 60 Compensation (USD) 2.75 4.50 5.38 4.50 5.00 5.00 5.75 5.00 4.40 4.40 4.40 4.00 4.00 # Excluded (Data Corrupt or Unrecorded) 1 0 0 0 1 1 0 1 1 1 1 0 0 # Excluded (Insufficient Cover Task Performance) 0 10 13 0 n/a n/a 0 n/a 0 9 9 0 (early boots) 0 (early boots) # Excluded (Insufficient TD Task Performance) 3 4 3 13 10 29 8 19 9 11 9 n/a n/a # Excluded (Insufficient 2AFC Task Performance) 15 4 2 8 21 6 8 3 11 5 11 1 1 # Excluded (Could Not See All Images) 0 3 4 4 1 4 2 1 6 8 6 0 1 # Excluded (Excessive “Focus Off” Events) 5 7 10 7 8 11 14 19 7 10 7 0 2 # Excluded (Previous participant in 3 3 7 3 1 3 8 2 9 11 9 1 2 176 similar experiment) # Excluded (Excessive Keypresses) 5 4 6 11 25 35 209 35 9 20 9 1 0 # Excluded (Did not understand or disregarded instructions) 9 16 17 4 7 6 12 6 17 25 17 0 3 # Excluded (Exceeded Experimental Duration Threshold) 0 0 3 0 0 1 0 0 0 2 0 0 0 Total Excluded 41 51 50 74 96 461 86 69 102 78 3 9 # “Held Out” (Unanalyzed) 41 13 31 19 28 23 19 18 12 46 12 8 13 Total Collected 202 184 189 222 239 600 224 135 268 210 71 82 Experimental Parameters Supplemental Table 2: Experimental parameters across conditions Attend Jiggle (0 Repetitions) Attend Jiggle (60 Repetitions) Attend Jiggle (96 Repetitions) 1-back Attend Stimuli Attend Relationships 2-back Attend Triplets Feedback Training Traditional TD Scrambled TD Non-famous Faces Famous Faces Cover Task Stimulus Duration (ms) n/a 800 800 800 800 800 800 800 800 800 800 800 800 Cover Task ISI (ms) n/a 200 200 200 200 200 1200 200 200 200 200 200 200 Cover Task Stimulus Repetitions n/a 60 96 60 60 60 60 60 2-20 60 60 60 60 TD Stimulus Duration (ms) 200 200 200 200 200 200 200 200 200 200 200 n/a n/a 177 TD ISI (ms) 40 40 40 40 40 40 40 40 40 40 40 n/a n/a # Elements per TD trial 7-11 7-11 7-11 7-11 7-11 7-11 7-11 7-11 7-11 12 12 n/a n/a # Targets per TD trial 1 1 1 1 1 1 1 1 1 1 1 n/a n/a PsiTurk version 2.1.2 2.1.2 2.1.2 3.0.0 3.0.0 3.0.0 3.0.0 3.0.0 2.2.4 2.2.3 2.2.3 3.0.0 3.0.0 Monte Carlo Simulations This section provides additional detail regarding the Monte Carlo simulations used to generate null distributions for the direct measures, and the bootstrapping procedure used to generate confidence intervals for observed data. For all simulations or resampling, the parameters were set to equate the simulated/resampled data with the observed data. The number of surrogate participants was equated with the experiment being analyzed (usually n=120). The number of simulations and bootstrap iterations performed was 15,000. The number of: stimuli (12), triplets (4), elements per triplet (3), and 2AFC trials (32) were the same for all experiments. Code is available on the Open Science Framework (https://osf.io/vtmpb). Determining whether the observed data was significantly different than chance required estimating an unbiased null hypothesis distribution. Our approach to generating such a distribution was to simulate results for surrogate participants across many surrogate experiments. Once the null distribution was generated, we then determined critical thresholds using our pre-registered hypotheses. We used one-tailed tests for the 2AFC and Creation simulations, given that we expected learning to be above chance. In addition to the critical thresholds for α=0.05, we also calculated the critical thresholds for α=0.01 & α=0.001. If the actual mean exceeded any of three levels of significance (p<0.05, p<0.01, and p<0.001), we reported the most unlikely to have occurred by chance. Otherwise, we reported only the measured value. 2AFC Simulation. An array of length 32 (i.e. number of trials) was populated with half zeros and half ones to represent correct left (0) or right (1) choices. This array was used as “ground truth” for all iterations. For each iteration, a surrogate experimental mean was calculated by first calculating a surrogate subject mean for each surrogate participant, and then taking the overall mean of all surrogate participants for that iteration. The mean for a surrogate participant was calculated by generating an array of 32 elements (i.e. number of trials) and randomly populating it with zeros and ones. This allowed us to calculate the “number 2AFC correct” for each surrogate participant. After all 15,000 iterations completed, we sorted the surrogate experimental means into ascending order to reveal the null distribution, calculating the overall chance experimental 178 mean and a critical threshold for “chance” performance, using an alpha of 0.05 (i.e. taking the value at 0.95 as the critical threshold). Creation Simulation. This simulation was used to calculate chance performance for both triplets created and foils created. An array of length 12 (i.e. number of stimuli) was populated with unique values from 0-11 (inclusive) to represent the stimulus identities. This array was used as “ground truth” for all iterations. Triplets were defined by splitting these 12 values into sets of three (i.e. “0”, “1”, “2” is triplet 1). For each iteration, a surrogate experimental mean was calculated. In each iteration, we looped through each fake participant and calculated the surrogate experimental mean by generating an array of 12 elements (i.e. number of stimuli) and randomly populated it with each value from 0-11. We then split the random 12 values into sets of three, allowing us to calculate the “number correct” for each surrogate participant. After all 15,000 iterations completed, we sorted the values in ascending order to create a null distribution. From this distribution, we calculated the overall chance experimental mean and a critical threshold for “chance” performance, using an alpha of 0.05 (i.e. taking the value at 0.95 as the critical threshold). Bootstrap Procedure. 95% confidence intervals were calculated using a bootstrap procedure. We ran multiple “surrogate experiments”. In each surrogate experiment, we sampled randomly (with replacement) from the actual experimental data (e.g. from the list of 2AFC percent-correct scores across participants) to generate a surrogate pool of 2AFC values, and computed the mean 2AFC performance for that surrogate experiment. Finally, we sorted the means from all the surrogate experiments in ascending order, employing the 2.5 and 97.5 percentiles as the boundaries of the 95% confidence interval. Counterbalancing procedure for stimulus sets. Counterbalancing was achieved by creating “mini-groups” of three stimulus sets, where the item positions were circularly shifted within a triplet across stimulus sets. Thus, subject X might be assigned the first stimulus set in a mini-group, and would see a triplet composed of the stimuli: (A1, A2, A3). Subject Y might be assigned the second stimulus set in the mini-group, and would see a triplet composed of the stimuli: (A2, A3, A1). Finally, subject Z, would see a triplet composed of: (A3, A1, A2). In this way, each stimulus image, A1 or A2 or A3, appears equally in the first, second and third position of a triplet. This circular shifting procedure ensures triplet position is uncorrelated with the properties of individual stimuli. 179 Supplemental Figure 2: Experimental paradigm for 1-back condition. Note the repeated image in the leftmost column. Self-assessment of performance. During the post-experiment questionnaire, we asked participants to assess their performance in each of the measures of learning after completing the experiment. They were separately asked, “How do you think you did in the {Target Detection, Group Choice, Triplet Creation} Task?”, and could answer with 1 of 4 options: “Very Well”, “Well”, “Poorly”, “Very Poorly”. Supplemental Results 180 Supplemental Table 3: Cover task performance, target detection accuracy, responses to questionnaire, and performance based upon responses. Attend Jiggle (0 Repetitions) Attend Jiggle (60 Repetitions) Attend Jiggle (96 Repetitions) 1-back Attend Stimuli Attend Relationships 2-back Attend Triplets Feedback Training Traditional TD Scrambled TD Non-famous Faces Famous Faces Experiment 1 1 1 2 2 2 2 2 2 3 3 4 4 Mean Cover Task Performance n/a 90% Jiggles Caught 90% Jiggles Caught 83% Repeats Caught n/a n/a 59% Repeats Caught n/a 3.07 attempts before 4 triplets created 92% Jiggles Caught 92% Jiggles Caught 90% Repeats Caught 90% Repeats Caught Mean Cover Task Response Time (ms) n/a 444 455 526 n/a n/a 742 n/a n/a 442 440 521 530 Mean TD Task Performance 90% Targets Caught 90% Targets Caught 89% Targets Caught 86% Targets Caught 89% Targets Caught 85% Targets Caught 89% Targets Caught 84% Targets Caught 86% Targets Caught 89% Targets Caught 91% Targets Caught n/a n/a Mean Target Response Time (ms) 424 410 414 443 430 422 434 426 417 433 439 n/a n/a Mean RTF1-3 (ms) 0.2 3.2 -1.1 -5.1 1.1 9.8 12.4 -4.7 -10.7 31.8 22.2 n/a n/a Mean 2AFC 51.4 50.2 51.7 54.0 57.1 62.3 66.9 66.9 81.3 51.9 48.9 63.3 59.3 Mean Triplet Creation 0.15 0.13 0.15 0.15 0.32 0.74 0.94 0.97 3.3 0.24 0.15 0.60 0.53 # Endorsing “Patterns Seen” before 2AFC 20 9 20 19 31 40 28 44 14 22 17 17 19 RTF1-3 for Pattern Seers (ms) -6.68 9.25 -9.74 -21.9 -4.09 18.0 23.3 4.03 -14.1 44.8 25.2 n/a n/a RTF1-3 for Pattern Unseers (ms) 1.61 2.70 0.64 -1.93 2.95 5.78 9.13 -9.82 -9.45 29.0 21.7 n/a n/a RTF1-3 Patterns Seen 907 (n.s., n.s.) 508 (n.s., n.s.) 905 (n.s., n.s.) 773 (n.s., n.s.) 1317 (n.s., n.s.) 1696 (n.s., n.s.) 1574 (n.s., n.s.) 1876 (n.s., n.s.) 228 (n.s., n.s.) 1324 (n.s., n.s.) 956 (n.s., n.s.) n/a n/a 181 Comparison (Mann-Whitney U, uncorrected, corrected) 2AFC for Pattern Seers 47.6 53.1 49.3 54.0 66.5 78.2 79.2 76.4 84.3 52.7 47.7 59.7 72.4 2AFC for Pattern Unseers 52.1 50.0 52.2 54.0 53.9 54.4 63.2 61.4 80.3 51.7 49.1 64.7 53.3 2AFC Patterns Seen Comparison (Mann-Whitney U, uncorrected, corrected) 809 (n.s., n.s.) 534 (n.s., n.s.) 863 (n.s., n.s.) 984 (n.s., n.s.) 1918 (p=0.0013, p=0.014) 2543 (p<0.001, p<0.001) 1857 (p<0.001, p=0.0046) 2366 (p<0.001, p=0.0017) 285 (n.s., n.s.) 1105 (n.s., n.s.) 857 (n.s., n.s.) 299 (n.s., n.s.) 593 (p=0.0012, p=0.016) Creation for Pattern Seers 0.20 0.44 0.0 0.053 0.55 1.58 1.79 1.34 3.43 0.32 0.12 0.41 1.26 Creation for Pattern Unseers 0.14 0.11 0.18 0.17 0.26 0.33 0.68 0.75 3.25 0.22 0.16 0.67 0.20 Creation Patterns Seen Comparison (Mann-Whitney U, uncorrected, corrected) 1096 (n.s., n.s.) 618 (p=0.040, n.s.) 860 (n.s., n.s.) 905 (n.s., n.s.) 1601 (n.s., n.s.) 2263 (p<0.001, p<0.001) 1806 (p<0.001, p=0.0036) 2112 (p=0.0083, n.s.) 291 (n.s., n.s.) 1042 (n.s., n.s.) 858 (n.s., n.s.) 343 (n.s., n.s.) 556 (p<0.001, p=0.011) # Pattern Seers who verbalized triplet structure 0 0 0 0 0 3 0 n/a n/a 0 0 0 0 # Pattern Seers who mention any structure 5 1 4 4 16 23 17 n/a n/a 6 5 9 7 # Endorsing “Knew there were 4 triplets” 14 11 8 27 11 23 30 41 43 14 9 9 13 RTF1-3 for 4 Triplet Endorsers (ms) -10.7 -6.43 -5.54 -27.3 -18.3 29.5 18.5 -0.77 -5.61 32.1 27.5 n/a n/a 2AFC for 4 Triplet Endorsers 52.6 46.7 50.2 52.6 58.5 67.1 66.8 73.0 84.7 51.8 48.3 56.6 56.9 182 Creation for 4 Triplet Endorsers 0.14 0.091 0.0 0.15 0.73 1.65 0.93 1.51 3.19 0.14 0.33 0.22 0.46 # Using Labels 26 30 31 39 38 43 68 39 29 30 34 n/a n/a RTF1-3 for Labellers (ms) 3.1 6.7 -3.4 -1.3 -0.56 26.3 12.7 -9.9 -17.0 35.1 18.7 n/a n/a RTF1-3 for No Labels -0.58 2.0 -0.29 -6.9 1.9 0.64 12.1 -2.3 -3.3 30.8 23.6 n/a n/a RTF1-3 Labels Used Comparison (Mann-Whitney U, uncorrected, corrected) 1313 (n.s., n.s.) 1461 (n.s., n.s.) 1357 (n.s., n.s.) 1775 (n.s., n.s.) 1532 (n.s., n.s.) 2044 (p=0.034, n.s.) 1817 (n.s., n.s.) 1499 (n.s., n.s.) 284 (n.s., n.s.) 1432 (n.s., n.s.) 1332 (n.s., n.s.) n/a n/a 2AFC for Labellers 51.5 49.4 52.1 54.2 56.6 72.8 72.5 79.4 87.1 52.5 47.3 n/a n/a 2AFC for No Labels 51.3 50.5 51.6 54.0 57.4 56.5 59.6 60.9 74.6 51.7 49.5 n/a n/a 2AFC Labels Used Comparison (Mann-Whitney U, uncorrected, corrected) 1198 (n.s., n.s.) 1142 (n.s., n.s.) 1349 (n.s., n.s.) 1510 (n.s., n.s.) 1568 (n.s., n.s.) 2341 (p<0.001, (p<0.001) 2444 (p<0.001, (p<0.001) 2367 (p<0.001, (p<0.001) 486 (p=0.033, n.s.) 1446 (n.s., n.s.) 1287 (n.s., n.s.) n/a n/a Creation for Labellers 0.15 0.23 0.29 0.13 0.29 1.35 1.32 1.77 3.24 0.20 0.15 n/a n/a Creation for No Labels 0.15 0.10 0.10 0.16 0.33 0.40 0.44 0.58 3.36 0.26 0.15 n/a n/a Creation Labels Used Comparison (Mann-Whitney U, uncorrected, corrected) 1276 (n.s., n.s.) 1542 (p=0.043, n.s.) 1524 (n.s., n.s.) 1470 (n.s., n.s.) 1548 (n.s., n.s.) 2226 (p<0.001, (p<0.001) 2412 (p<0.001, (p<0.001) 2303 (p<0.001, (p<0.001) 349 (n.s., n.s.) 1335 (n.s., n.s.) 1485 (n.s., n.s.) n/a n/a Jaccard Distance b/w Pattern Seeing & Labelling 0.82 0.95 0.79 0.84 0.79 0.64 0.77 0.72 0.84 0.89 0.79 n/a n/a # Rated TD performance as “Very Well” or “Well” 86 81 81 89 73 87 80 80 33 85 91 n/a n/a RTF1-3 for Well+ Endorsers (ms) 1.66 2.05 5.41 -2.19 -2.89 10.5 15.1 0.66 -12.1 31.3 24.3 n/a n/a 183 # Rated 2AFC performance as “Very Well” or “Well” 49 58 53 72 58 80 87 81 42 59 57 47 47 2AFC for Well+ Endorsers 50.5 50.6 50.3 52.3 59.9 66.7 67.5 67.9 83.7 52.4 51.2 64.3 59.7 # Rated Creation performance as “Very Well” or “Well” 42 30 28 59 36 60 66 63 40 37 29 34 34 Creation for Well+ Endorsers 0.21 0.20 0.21 0.19 0.58 1.05 1.35 1.27 3.70 0.43 0.31 0.82 0.82 Supplemental Figure 3: Triplet creation by triplet type. "Foil Triplet" refers to a sequence shown during the 2AFC, but not during the initial exposure or TD task. "Real triplet" indicates a triplet that was shown during the initial exposure and TD task. 184 Supplemental Figure 4: Relationships between response time facilitation and direct measures (2AFC and Creation). (a) Response time facilitation between first and third triplet positions and its relationship with 2AFC performance. (b) Response time facilitation between first and third triplet positions and its relationship with triplet creation performance. Black dots indicate means of distributions in both dimensions. Black line segments indicate 95% confidence intervals. Supplemental Figure 5: Response time facilitation between triplet positions 1 & 3 over the course of the TD task (split by quartile). 185 Supplemental Figure 6: Response time effects in modified TD conditions, across quartile of TD task. (a) Response times by triplet position, by TD quartile. (b) RT facilitation between first and third triplet positions, by quartile. Supplemental Discussion Motoric learning might be conflated with statistical learning in some paradigms Most learning occurs over time, so temporal structure is critical across all domains to-be-learned. In contrast to SRT paradigms (Nissen & Bullemer, 1987), the underlying neural mechanisms of learning the temporal structure of the stimuli occurring in the world are dissociable from any motor action (Olson & Chun, 2001). Thus, we designed our experiments to specifically focus on non-motor learning in the visual modality. Factors that affect visual statistical learning Our findings have revealed several areas that seem to affect associative learning similarly, including: method of engagement with regularities, method of test, and stimulus choice. We suggest that, specific to VSL, many experimental design decisions can influence the extent to which VSL occurs. Some of the clearest are: 1) Method of engagement (i.e. cover task during exposure phase) a. This includes the instructions provided to the participant before the exposure phase 186 b. Attention to temporal aspects > Attention to stimulus-specific aspects i. Attention to both is likely additive 2) Method of test a. Target Detection tasks with short streams have the strongest effects of the RT speeding by stream position b. The more segmentation in the exposure phase (or conversely the TD task), the higher the probability of detection of structure i. Unproven whether this correlates with explicit detection, but seems likely c. Explicit knowledge before test could affect performance (e.g when participants are explicitly told about the structure before 2AFC (e.g. Vickery et al., 2018)). 3) Stimulus choice a. Verbalizable labels > Unverbalizable labels b. The effects of familiarity require further investigation and pre-existing associations between stimuli in general Some design decisions that seem not to have a large effect on VSL: 1) Number of repetitions a. This finding may be specific to the Jiggle cover task. Future work should investigate how number of repetitions of the statistical structure interacts with choice of cover task. Given the work on item-specific vs relational vs both processing (Hunt & Einstein, 1981), it is reasonable to assume that participants in a condition which requires either item-specific or relational processing could benefit by then participating in a condition which required the other. In the original Einstein & Hunt studies, they showed that repetition of the same processing type did not improve memory. Therefore, I hypothesize the same thing could be occurring in the standard statistical learning paradigm (i.e. that item-specific processing many times leads to a plateauing learning effect). One could, alternatively, run a condition with many fewer repetitions, but with a cover task that switches. For example, do the Jiggle task for 5 repetitions of each triplet and then a 2-back for 5 repetitions and compare learning. I hypothesize that learning would be superior in this case than either the “Jiggle” or “2-back” conditions, individually, with a higher number of repetitions. This hypothesis is in direct contrast to our findings in Exp. 1, which was that repetition seemed to have no effect on statistical learning. I interpret this null effect as a consequence of the method of engagement with the regularities. Since the Jiggle Task does not require any item-specific processing (simply detect image movement) and certainly no relational elements, it’s likely that the amount of repetition was irrelevant. However, if the cover task was one which did require 187 relational processing (such as the 2-back), then I hypothesize that repetition would have a positive effect on the direct measures of learning. Implicit learning is affected by the probabilistic structure Performance in implicit learning paradigms are affected by difficulty of the statistical structure. Just as with the standard research paradigm when transition probabilities are <1 (e.g. Siegelman et al., 2017), the success rates in AGL tasks are correlated with the difficulty of the grammar (Christopher M. Conway, 2020). This could be support for a shared underlying learning mechanism or an indication that both processes are affected by increased cognitive load. Regardless, either interpretation could be taken as evidence against the automaticity view, since it is assumed to not be affected by other cognitive processing. Is generalizability a defining characteristic of statistical learning? One characteristic of statistical learning that is debated is the extent to which the regularities that are statistically learned become generalizable. Some argue that “transfer learning” from the statistical regularities is possible (Aslin & Newport, 2012; Aslin, 2017), thus enabling prediction of novel associations, while others believe the evidence for generalization is limited (Frost et al., 2015). Goldilocks, but for Adults Infants may only demonstrate VSL when the novelty of the stimuli is in a “sweet spot” between totally novel and totally well-known (Kidd et al., 2012). It is possible a similar phenomenon is occurring in adults. For example, novel and difficult-to-name stimuli (such as the fractals) may be difficult to associate. On the other extreme, for extremely well-known stimuli (e.g. famous faces), it may be difficult to associate. Stimuli that are processed fluently, yet lack a strong schematic association, might permit the most rapid association. This possibility could explain the results from Experiment 4. We found that non-famous faces were learned better than famous faces, against our hypotheses. Perhaps famous faces are too well known to permit superior VSL to non-famous faces. Similarly, both novel and famous faces were learned better than abstract, colorful fractals (Exp. 1, “1-back”), suggesting that fractals could be more extreme on the “too novel” side of the inverse-U of learning. These findings occurred when cover tasks were equated, so the only independent variable that was different was the choice of stimulus. Applications to pedagogy 188 Understanding regularity learning may have pedagogical implications. Previous work has shown a positive correlation between SL and reading abilities (Arciuli & Simpson, 2012). This finding can be interpreted in multiple ways. One possibility is that a regularity learning ability underlies both SL and reading. An alternative explanation is that memory strength and processing fluency for the stimuli to-be-associated might influence the efficacy of both SL and reading. While it is difficult to dissociate the learning which results from a regularity learning process and the influence of individual stimuli, future work should investigate this distinction. We suggest that focus should be made on equating the method of engagement (Exp. 2) and stimulus choice (Exp. 4) to tease apart differences in SL and reading abilities. Connections between automatic statistical learning and attentional capture How does the “automatic detection of regularities” differ from the attentional capture literature (Le Pelley et al., 2016)? Le Pelley and colleagues highlight the reciprocal nature of attention and learning, stating that attentional capture leads to an increased rate of learning. The same reciprocal process could be happening in statistical learning, but this has not yet been investigated from an “attentional capture” perspective. Priming in TD task Work assessing statistical learning via indirect measures likely reflect a faciliatory effect from similar processes to working memory (e.g. Batterink, 2017). This type of learning is likely ephemeral, and is likely to be quickly forgotten. This phenomenon should not be confused with learning that is durable enough to be tested after a delay and can be assessed using multiple measures. One important caveat to our findings is how they apply to longer TD test streams. TD implementations vary dramatically in test stream length (7-48 elements) and our implementation is on the shorter end of the range. We limited stream length to ensure there were no repetitions of images in any TD stream, to avoid any potential short-term priming effects. However, a longer test stream, with target repetitions, may experience an interaction between short-term priming and the speeding effect seen at the start of a trial or after a target is presented. We suggest that such measurements might reveal transient priming rather than a learning effect that will persist in long-term memory. Forced-choice performance is similar to other visual statistical learning studies We observed 2AFC percormance in-line with other VSL studies, including where the participants had previously participated in a statistical learning experiment (Bogaerts et al., 2016). 189 Could the weak learning in Experiments 1 & 3 be explained by a lack of attention during the initial exposure phase? One might suppose that the weak 2AFC performance in Experiments 1 & 3 could reflect a lack of attention to the stimuli. However, it is unlikely that participants in our experiments performed poorly due to lack of attention: we stringently selected participants for task compliance; they detected motion rapidly (mean RTs < 460ms) and accurately (>89%); and the participant pool is clearly able to learn, as shown in Experiment 2 (Supp. Table 3). The most likely explanation of the poor 2AFC performance in Experiments 1 & 3 is the method of exposure to the triplet regularities. Type of participant may influence SL, but future research is required Shapes data show equivalent learning online and in-lab (Suppl. Table 3), but in-lab performance was quantitatively better. How much of the excellent learning that is published on visual statistical learning is a byproduct of particularly motivated and intelligent subjects being used on college campuses? Other studies (Frost) show that nearly 1/3 of participants demonstrate no statistical learning. A fourth possibility on why there was no VSL difference between famous and non-famous stimuli The relevance of a potential association may limit or facilitate associative and VSL. Though Schwartz & Yovel (2016) found a memory benefit for famous relative to non-famous faces when associating a name or occupation, they found no memory benefit when the face was paired with a meaningless image (e.g. a random orthographic symbol). This suggests that there is no benefit of famousness in associating arbitrary stimuli, so the potential associations need to be plausibly related (e.g. Bein et al, 2019). Importance of documentation In addition, the details of experimental design and data processing become especially important for interpretation; we echo calls for documentation and sharing of experiments and their data (e.g. Stall et al., 2019). Our pre-registered experiments, code and data are available at https://osf.io/vtmpb. 190 Learning in other species Though VSL may not be wholly automatic, it does appear that statistical learning might be qualitatively different in humans relative to other living beings. For example, humans appear to easily extract regularities between items, whereas honey bees appear to automatically extract only stimulus frequencies (Avarguès-Weber et al., 2020). While this does not provide clear evidence of human VSL automaticity, it does suggest that co-occurrence statistics may be extracted more automatically in humans than in other species.

Edited by:

Abstract

Extracted data

Edited by:

Abstract

Extracted data

Related items

Related items