Proceedings of the Annual Meeting of the Cognitive Science Society

Existing models of statistical learning involve computation ofconditional probabilities over discrete, categorical items in asequence. We propose an alternative view that learning occursthrough a process of tracking changes along physicaldimensions from one stimulus to the next within a “perceptualsimilarity space.” To test this alternative, we examined asituation where it is difficult or impossible to label stimuli inreal time, and where the two assumptions lead to conflictinghypotheses. We conducted two experiments in which humanparticipants passively listened to a familiarization sequence offrequency-modulated tones and were then asked to makefamiliarity judgments on a series of test bigrams. Behavioralresults were broadly consistent with a conceptualization oflearning as tracking trajectories through perceptual similarityspace. We also trained a neural network that codes stimuli asvalues along two continuous dimensions to predict the nextstimulus given the current stimulus, and show that it capturedkey features of the human data.