Figure 2. Cross-temporal classification performance reveals dynamic coding.
Cross-temporal classification of 1st stimulus identity (left panel), 2nd stimulus identity (middle panel) and network response (right panel). Row i and column j of each matrix indicates the accuracy of a classifier, trained on population activity data at time i, in guessing a specific task feature using population activity data at time j (training and decoding data are always separate). While the network reliably encodes information about stimulus identity right until the onset of the response period (as shown by high accuracy values along the diagonal in left and middle panel), this information is stored with a highly dynamic encoding (as shown by low cross-classification accuracy across successive periods, i.e., ‘bottlenecks’ with high accuracy on the diagonal but low accuracy away from the diagonal). Note that in both left and middle panels, stimulus identity information decreases greatly at the onset of the response period, reflecting a shift to a from stimulus-specific to response-specific encoding (see also Figure 3).