Suitably chosen intermediate ConvRNN circuits provide consistent predictions of primate ventral stream neural dynamics. (a) The -axis indicates the median across neurons of the explained variance between predictions and ground-truth responses on held-out images divided by the square root of the internal consistencies of the neurons, defined in section A.6.3. Error bars indicates the s.e.m. across neurons ( for V4, for pIT, for cIT/aIT) averaged across 10 ms time bins ( each for the Early and Late designations). As can be seen, the intermediate-depth feedforward BaseNet model (first bars) is a poor predictor of the subset of late responses that are beyond the feedforward pass, but certain types of ConvRNN circuits (such as RGC Median, UGRNN, and GRU) added to the BaseNet are overall best predictive across visual areas at late time points (Wilcoxon test, with Bonferroni correction with feedforward BaseNet, for each visual area). See Figure S6 for the full time courses at the resolution of 10 ms bins. (b) For each ConvRNN circuit, we compare the average neural predictivity (averaged per neuron across early and late timepoints) averaged across areas, to the OST consistency. The ConvRNNs that have the best average neural predictivity also best match the OST consistency (RGC Median, UGRNN, and GRU).
