Regularizing Neural Networks by Penalizing Confident Output Distributions

Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, Geoffrey Hinton

TL;DR: We show that penalizing low entropy output distributions, which has been shown to improve exploration in reinforcement learning, acts as a strong regularizer in supervised learning.

Abstract: We propose regularizing neural networks by penalizing low entropy output distributions. We show that penalizing low entropy output distributions, which has been shown to improve exploration in reinforcement learning, acts as a strong regularizer in supervised learning. We connect our confidence penalty to label smoothing through the direction of the KL divergence between networks output distribution and the uniform distribution. We exhaustively evaluate our proposed confidence penalty and label smoothing (uniform and unigram) on 6 common benchmarks: image classification (MNIST and Cifar-10), language modeling (Penn Treebank), machine translation (WMT'14 English-to-German), and speech recognition (TIMIT and WSJ). We find that both label smoothing and our confidence penalty improve state-of-the-art models across benchmarks without modifying existing hyper-parameters.

Conflicts: google.com

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/arxiv:1701.06548/code)