The training accuracy for neural networks, colored from black (random chance) to red (high accuracy). Overlaid in white dashed lines are the theoretical predictions showing the boundary between trainable and untrainable networks. (a) Networks with no dropout. (b)-(d) Networks with dropout rates of 0.01, 0.02, 0.06 respectively. This research explores whether theoretical calculations can replace large hyperparameter searches. For more details, read “Deep Information Propagation” (S. S. Schoenholz, J. Gilmer, S. Ganguli, J. Sohl-Dickstein, submitted to ICLR 2017). |