Figure 1. Delayed nonmatch-to-sample task.
(A) (top): task description. The network is exposed to two successive stimuli, with an intervening delay. The task is to produce output −1 if the two stimuli were identical (AA or BB), or 1 if they were different (AB or BA); the output of the network is simply the activity of one arbitrarily chosen ‘output’ neuron, averaged over the last 200 ms of the trial. (B) (bottom left): time course of trial error (mean absolute difference between output neuron response and correct response over the last 200 ms of each trial) during learning over 10000 trials (dark curve: median over 20 runs; gray area: inter-quartile range). The solid vertical line indicates the median number of trials needed to reach the criterion of 95% ‘correct’ trials (trial error <1) over 100 successive trials (843 trials); dotted vertical lines indicate the inter-quartile range (692–1125 trials). Performance (i.e., magnitude of the response error) continues to improve after reaching criterion and reaches a low, stable residual asymptote. ( (bottom right): Activities of 6 different neurons, including the output neuron (thick black line), for two stimulus combinations, before training (left) and after training (right). Note that neural traces remain highly dynamical even after training.