Analyzing noise in autoencoders and deep networks

TL;DR

This paper introduces a unified noise injection framework for autoencoders and deep networks, demonstrating that various regularization techniques can be interpreted within this framework and that noise improves representation learning and denoising performance.

Contribution

The paper extends denoising autoencoders by injecting noise before nonlinearities and at hidden units, unifying multiple regularization methods under a single framework, and showing practical benefits in deep learning.

Findings

Noisy autoencoders outperform denoising autoencoders in denoising tasks.

Injected noise improves performance on MNIST and CIFAR-10 datasets.

Different types of noise enhance deep network representations through sparsity and decorrelation.

Abstract

Autoencoders have emerged as a useful framework for unsupervised learning of internal representations, and a wide variety of apparently conceptually disparate regularization techniques have been proposed to generate useful features. Here we extend existing denoising autoencoders to additionally inject noise before the nonlinearity, and at the hidden unit activations. We show that a wide variety of previous methods, including denoising, contractive, and sparse autoencoders, as well as dropout can be interpreted using this framework. This noise injection framework reaps practical benefits by providing a unified strategy to develop new internal representations by designing the nature of the injected noise. We show that noisy autoencoders outperform denoising autoencoders at the very task of denoising, and are competitive with other single-layer techniques on MNIST, and CIFAR-10. We also show that types of noise other than dropout improve performance in a deep network through sparsifying, decorrelating, and spreading information across representations.