eLife assessment
This work provides a valuable analysis of the effect of two commonly used hyperparameters, noise amplitude and firing rate regularization, on the representations of relevant and irrelevant stimuli in trained recurrent neural networks (RNNs). The results suggest an interesting interpretation of prefrontal cortex (PFC) dynamics, based on comparisons to previously published data from the same lab, in terms of decreasing metabolic cost during learning. The evidence indicating that the mechanisms identified in the RNNs are the same ones operating in PFC was considered incomplete, but could potentially be bolstered by additional analyses and appropriate revisions.