Regularizing Neural Networks by Penalizing Confident Output Distributions

We systematically explore regularizing neural networks by penalizing low\nentropy output distributions. We show that penalizing low entropy output\ndistributions, which has been shown to improve exploration in reinforcement\nlearning, acts as a strong regularizer in supervised learning. Furthermore, we\nconnect a maximum entropy based confidence penalty to label smoothing through\nthe direction of the KL divergence. We exhaustively evaluate the proposed\nconfidence penalty and label smoothing on 6 common benchmarks: image\nclassification (MNIST and Cifar-10), language modeling (Penn Treebank), machine\ntranslation (WMT'14 English-to-German), and speech recognition (TIMIT and WSJ).\nWe find that both label smoothing and the confidence penalty improve\nstate-of-the-art models across benchmarks without modifying existing\nhyperparameters, suggesting the wide applicability of these regularizers.\n