The figures below show that the dynamic population coding of information is due to different neurons coding the same information at different points in time (as opposed to having the same subset of neurons coding the information at all times, and the dynamics being do to these neurons going through adaptation-like effects). To show that different neurons have the match/nonmatch information at different time periods, we have done an analysis in which we eliminate the 256 most selective neurons at one time period using the training data and then we have trained and tested the classifier using the remaining neurons. Neuron selectivity was determined using an ANOVA and the 256 neurons with the smallest p-values were eliminated (i.e., 750 - 256 = 494 neurons were used for the feature task and 600 - 256 = 344 neurons were used for the spatial task). As can be seen by the green diagonal band in the figure, decoding performance goes down dramatically when the time period used to eliminate the neurons is the same time period used to train and test the classifier indicating that removing good neurons decreasing the classifier performance (which is similar to the results shown in Figs. S2b and S3e). However, what is most interesting is that when neurons are eliminated at one time period but the training and testing is done at a different time period, the results are only slightly lower than when random neurons are eliminated, which indicates that different neurons have the match/nonmatch information at different points in time. Results from the feature task are shown in the upper figure and the results from spatial task are shown in the lower figure.
