Architecture of RNNs and simulation results.
(A) All networks consist of 3 layers of artificial units: the input, hidden and output layers. For both RNN1 and RNN2, the input layer contains 20 units including 15 orientation-tuned units (red) and 5 cue units (retro-cue and task cue, orange and yellow). The hidden layer consists of three modules of 200 recurrent units with short-term synaptic plasticity (STSP), further divided into 80% excitatory (black) and 20% inhibitory (white). Connectivity within each module (black arrow) is denser compared to between modules (red and green arrows), which only occur between excitatory units. Only excitatory units in Module 1 receive projections from the input layer and only excitatory units in Module 3 project to the output units. For RNN1, networks output (0,1) or (1,0) through the 2 units in the output layer to indicate responses. For RNN2, the network output (0,1) or (1,0) to report the category to which the cued orientation belonged in the categorization task, or (0,0) in the maintenance task (blue units). Importantly, the models also output the orientation itself through 15 additional orientation-tuned units (red). (B) Difference in stimulus decoding between tasks in RNN1 (upper panel) and RNN2 (lower panel). Results were averaged across the delay period. Positive difference indicates higher decoding accuracy for categorization, and negative difference indicates higher decoding accuracy for maintenance. The inset above illustrates stimulus difference in human fMRI results during late epoch in Experiment 1 to provide a reference for expected patterns in RNNs. (C) Average abstract category information across the delay period for RNN1 (upper panel) and RNN2 (lower panel). The inset above illustrates abstract category representation in human fMRI. Error bars represent ± SEM. Black asterisks denote FDR-corrected significance, *: p < 0.05; **: p < 0.01; ***: p < 0.001.