A New Model of Statistical Learning: Trajectories Through Perceptual Similarity Space

Abstract

Existing models of statistical learning involve computation of conditional probabilities over discrete, categorical items in a sequence. We propose an alternative view that learning occurs through a process of tracking changes along physical dimensions from one stimulus to the next within a “perceptual similarity space.” To test this alternative, we examined a situation where it is difficult to categorize stimuli, and where the two assumptions lead to different predictions. We conducted two experiments in which participants passively listened to a familiarization sequence of frequency-modulated tones and were then asked to make familiarity judgments on a series of test bigrams. Results were broadly consistent with a conceptualization of learning as tracking trajectories through perceptual similarity space. We also trained a neural network that codes stimuli as values along two continuous dimensions to predict the next stimulus given the current stimulus, and show that it captured key features of the human data.


Back to Table of Contents