Associative Processes in Statistical Learning: Paradoxical Predictions of the Past

The ability to process sequences of input and extract regularity across the distribution of input is fundamental for making predictions from the observed past to the future. Prediction is rooted in the extraction of both frequency- and conditional statistics from the distribution of inputs. For example, an animal hunting for food may consistently return to a particular area to hunt if relative to all other areas visited, that area has the highest frequency of prey. In contrast, humans asked to predict the next word in a sentence must make a prediction based upon higher-order regularities rather than simple frequency statistics (the most frequent word in the English language is the). The Serial Reaction Time (SRT) task, a model for studying sequential behavior, is used to quantify sensitivity to sequential constraints present in structured environments (Nissen & Bullemer, 1987). The SRT task requires Ss to make a unique response to each individually presented element from a sequence of elements. The statistics of SRT sequences, such as the relative frequency of elements and simple pairwise associations between elements, can be controlled to create dependencies that can only be predicted by learning higher-order associations. Sensitivity to the sequential constraints present in the structured input is demonstrated through differences in reaction time to elements that are, and are not, predictable based upon the statistics of the input environment. Sensitivity to statistical regularity in the environment is also a critical dimension of various episodic learning methodologies. Graded associations have been demonstrated among elements extending in both forward and backward directions in episodic memory tasks, and are suggested to reflect a gradient of the underlying structural relationships among the study elements. Graded associations are beneficial to the extent that they increase the probability of recalling sequence elements.However, unlike free and serial recall tasks, backward associations, and remote associations in general, are anti-predictive in the SRT task. The formation of associations beyond the immediately predictive element in prediction tasks could be suggestive of a ubiquitous underlying associative mechanism, which universally gives rise to graded contiguity effects, regardless of the specifc application (Howard, Jing, Rao, Provyn, & Datey, 2009). The following experiment employed a probabilistic SRT task to quantify sensitivity to immediately backward, backward-remote, and forward-remote associations. Ss were presented sequences of elements probabilistically sampled from an underlying ring-structure, with the dependent measure Ss\u27 reaction time to elements that either followed, or deviated from, the structure. Results from the SRT task indicated that Ss demonstrated a robust backward association, as well as evidence for forward-graded associations. Moreover, in an explicit test of sequence knowledge, while Ss did not generate the probabilistic statistics from the structured learning environment, Ss did generate a statistically signifcant amount of backward-transitions, relative to other remote-backward transitions. The graded associations that were formed beyond the immediately predictive element in the prediction task provide evidence that a similar mechanism that mediates episodic learning may also mediate statistical learning. Backward and graded associations may be explained by a ubiquitous underlying associative mechanism, which universally gives rise to graded contiguity effects, regardless of the specific application