Fig. 3: Computational model of sensory prediction errors in perirhinal cortex.
From: Perirhinal cortex learns a predictive map of the task environment
a An autoencoder with three layers (input, hidden, and output) was trained to represent the input. The input consisted of two independent stimulus variables: direction of motion (red) and speed (blue). A linear classifier was trained to decode direction (red) and speed (blue) by reading out the difference between the reconstructed output and the input (dotted line). Sparsity in the hidden layer was imposed by adding an L1-norm term on the loss function. b Decoding performance of direction (red, left) and speed (right, blue) as a function of training epoch for the linear classifier reading out from familiarity activity. Similar to experimental results in Prh, decoding performance of direction decreases, whereas decoding performance for speed increases throughout training. Error bars correspond to SEM across independent simulations (n = 100). See also Supplementary Fig. 8.
