Robustness.

A: Noise robustness heat maps for sparse RNNs. The horizontal axis indicates the level of dynamic noise, defined as the standard deviation of the dynamic noise divided by the mean of the activity in the RNN, and the vertical axis indicates the standard deviation of input noise. Sparse RNNs maintain high accuracy over a wide range of noises levels. # epochs = 150. B: Output of an example sparse RNN over time. Model is trained for 2τ, and input is on for τ. N = 300, f = 0.05, α = 1, # epochs = 500. C: Performance of the E/I sparse RNNs as a function of the E/I ratio. N = NE + NI = 100, f = 0.1, α = 1, # epochs = 200.