A Computational Model which Learns to Selectively Attend in Category Learning

Shepard et al. (1961) made empirical and theoretical investigation of the difficulties of different kinds of classifications using both learning and memory tasks. As the difficulty rank mirrors the number of feature dimensions relevant to the category, later researchers took it as evidence that category learning includes learning how to selectively attend to only useful features, i.e. learning to optimally allocate the attention to those dimensions relative to the category (Rosch and Mervis, 1975). We built a recurrent neural network model that sequentially attended to individual features. Only one feature is explicitly available at one time (as in Rehder and Hoffman's eye tracking settings (Render and Hoffman, 2003)) and previous information is represented implicitly in the network. The probabilities of eye movement from one feature to the next is kept as a fixation transition table. The fixations started randomly without much bias on any particular feature or any movement. The network learned the relevant feature(s) and did the classification by sequentially attending to these features. The rank of the learning time qualitatively matched the difficulty of the categories