A neural network model of how category learning alters perceptual similarity

Decades of research have shown that how we categorize objects changes how we perceive them. For example, category learning can increase the perceptual similarity of within-category items and decrease the similarity of between-category items (Goldstone and Hendrickson, 2010). However, the mechanism for this phenomenon (commonly called Categorical Perception or CP) is under debate. Based on previous work showing that CP is reduced when attention is directed away from the stimulus (Brouwer and Heeger, 2013), here we postulate that categorical perception occurs through category-learning steered Feature-Based Attention (FBA). In this theory, category learning alters perception by steering FBA to change the gain of visual cortex neurons responding to specific feature values. We test this hypothesis utilizing two category structures: Rule-Based (RB) and Information-Integration (II). In RB structures, the optimal strategy is to make binary decisions along feature dimensions and performance is increased if attention is targeted to specific feature values. In II structures, information from multiple feature dimensions must be combined before a decision can be made and performance is decreased if attention is targeted to specific feature values. The theory predicts that RB structures will cause greater changes in perception than II structures. Subjects (n=20) were divided evenly between RB and II conditions and underwent same-different testing before and after category learning. We found that the RB condition caused significantly more changes in same-different performance than the II condition. We then implement a neural network model that learns to apply feature-specific feedback (gain) modulation during category learning. We show that the feedback connections in the model, which was previously developed to explain how category learning induces transfer of visual perceptual learning, enable the network to show the same behavior patterns as human participants. This work provides strong behavioral and computational evidence for feature-based attention being the mechanism for categorical perception.