Fig. 4: Using minimodels to understand visual invariances.

a Visual textures from 16 categories were shown to the mice in addition to the natural images. b Decoding accuracy of texture class on test images (n = 4 mice). c Trial-averaged responses of two example neurons (i and j) to the test images from two different texture classes (10 trials per test image). d Comparing the category variance of model neurons and recorded neurons. Pearson correlation (r) and p value of two-sided test reported. e Mean category variance of model features after each successive operation. f Category variance of the model prediction vs pooling diameter in the readout layer. Pearson correlation (r) and p value of two-sided test reported. g Same as (f) for the input diversity, which is defined as the mean correlation between conv2 channels with positive w_c weights. Pearson correlation (r) and p value of two-sided test reported.