Direct associations or internal transformations? Exploring the mechanisms underlying sequential learning behavior

Abstract We evaluate two broad classes of cognitive mechanisms that might support the learning of sequential patterns. According to the first, learning is based on the gradual accumulation of direct associations between events based on simple conditioning principles. The other view describes learning as the process of inducing the transformational structure that defines the material. Each of these learning mechanisms predicts differences in the rate of acquisition for differently organized sequences. Across a set of empirical studies, we compare the predictions of each class of model with the behavior of human subjects. We find that learning mechanisms based on transformations of an internal state, such as recurrent network architectures (e.g., Keywords: Sequence learning; Skill acquisition and learning; Learning constraints; Simple recurrent networks; Associative learing The ability to learn about the stream of events we experience allows us to perceive melody and rhythm in music, to coordinate the movement of our bodies, and to comprehend and produce utterances; it forms the basis of our ability to predict and anticipate. Despite the ubiquity of sequential learning in our mental lives, an understanding of the mechanisms which underlie this ability has proven elusive. Over the last hundred years, the field has offered at least two major perspectives concerning the mechanisms supporting this ability. The first (which we will call the ''Associationist'' or ''Behaviorist'' view) describes learning as the process of incrementally acquiring associative relationships between events (stimuli or responses) that are repeatedly paired. A defining feature of this paradigm is the claim that behavior can be characterized without reference to internal mental processes or representations A second view, offered by work following from the ''cognitive revolution,'' highlights the role that the transformations of intermediate representations play in cognitive processing. Broadly speaking, transformational systems refer to general computational systems whose operation involves the iterative modification of internal representations. By this account, processes which apply structured transformations to internal mental states interact with feedback from the environment in order to control sequential behavior. For example, a cornerstone of contemporary linguistics is the idea that language depends on a set of structured internal representations (i.e., a universal grammar) which undergo a series of symbolic transformations (i.e., transformational grammar) before being realized as observable behavior Transformations are often more flexible than simple associations. For example, systems that depend on symbolic rewrite rules can apply the same production when certain conditions are met, irrespective of surrounding context (i.e., a rewrite rule such as xABy fi xBAy, which simply flips the order of elements A and B, applies equally well to strings like FXAB-TYU and DTABRRR due to the presence of the common AB pattern). This flexibility allows generalization between structurally similar inputs and outputs. In addition, while the selection of which transformation to apply at any given point may be probabilistic, the rules used in transformations tend to regularize their output since only a single mapping can be applied at any time. Critically, transformational systems depend on an appropriately structure intermediate ''state'' or representation which is modified through additional processing (in contrast to the associationist account which learns direct relations between observed behavioral elements). The importance of transformations as a general cognitive mechanism extends beyond strictly linguistic domains such as transformational grammars for which they are most easily identified. For example, theories of perceptual and conceptual similarity have been proposed which assume that the similarity of two entities is inversely proportional to the number of operations required to transform one entity so as to be identical to the other To foreshadow, our results show that, at least on shorter time scales, human sequence learning appears more consistent with a process based on simple, direct associations. This conclusion is at odds with recent theoretical accounts which have argued for domain-general sequential learning processes based on transformations of an internal state (such as the simple recurrent network, Associations versus transformations: Two representative models We begin our analysis by considering two models that are representative of the broad theoretical classes just described. The first, called the linear associative shift-register (LASR), is a simple associative learning system inspired by classic models of error-driven associative learning from the associationist tradition . The second is the simple recurrent network (SRN) first proposed by The LASR model LASR is a model of sequential processing based on direct stimulus-stimulus or stimulusresponse mappings. The basic architecture of the model is shown in The model is organized around two simple principles. First, LASR assumes a simple shift-register memory for past events (see 1 Thus, the most recent event is always located in the right-most slot of the register (see Second, the simple shift-register memory mechanism is directly associated with output units called detectors without mediation by any internal representations or processing (se