Performance.

A: Accuracy for various combinations of N and P. Colors represent accuracy, grouped in bins spanning 3%. Lines are fitted to every group of binned accuracy values. f = 0.1, # epochs = 200 and number of trained networks (n): n = 4. B: Binary classification performance of sparse RNNs of various sizes. Each RNN classifies P = αfN2 inputs. f = 0.1, α = 1, # epochs = 500, n = 4. C: Sparse RNN performance as a function of the readout sparsity for RNNs with two different levels of recurrent sparsity. N = 100, α = 0.5, # epochs = 500, n = 6. D: Accuracy of sparse RNNs as a function of α = P/(fN2). Runs were truncated at 99% accuracy. N = 100, f = 0.05, # epochs = 5000, n = 4.