Deep Networks with Stochastic Depth

TL;DR

Stochastic depth is a training method for very deep residual networks that randomly drops layers during training, enabling faster training and improved test accuracy even at depths exceeding 1200 layers.

Contribution

This paper introduces stochastic depth, a novel training procedure that allows deep residual networks to be trained efficiently by randomly bypassing layers during training.

Findings

Reduces training time significantly.

Improves test error across multiple datasets.

Enables training of networks with over 1200 layers.

Abstract

Very deep convolutional networks with hundreds of layers have led to significant reductions in error on competitive benchmarks. Although the unmatched expressiveness of the many layers can be highly desirable at test time, training very deep networks comes with its own set of challenges. The gradients can vanish, the forward flow often diminishes, and the training time can be painfully slow. To address these problems, we propose stochastic depth, a training procedure that enables the seemingly contradictory setup to train short networks and use deep networks at test time. We start with very deep networks but during training, for each mini-batch, randomly drop a subset of layers and bypass them with the identity function. This simple approach complements the recent success of residual networks. It reduces training time substantially and improves the test error significantly on almost all data sets that we used for evaluation. With stochastic depth we can increase the depth of residual networks even beyond 1200 layers and still yield meaningful improvements in test error (4.91% on CIFAR-10).