Extracting and composing robust features with denoising autoencoders

TLDR

Deep generative or discriminative models can be trained more effectively by an initial unsupervised step that learns useful intermediate representations. The study proposes a training principle that makes representations robust to partial input corruption. The authors train denoising autoencoders that can be stacked to initialize deep architectures, motivated by manifold learning, information theory, or generative modeling. Experiments demonstrate that corrupting inputs during autoencoder training improves performance on pattern classification benchmarks.

Abstract

Previous work has shown that the difficulties in learning deep generative or discriminative models can be overcome by an initial unsupervised learning step that maps inputs to useful intermediate representations. We introduce and motivate a new training principle for unsupervised learning of a representation based on the idea of making the learned representations robust to partial corruption of the input pattern. This approach can be used to train autoencoders, and these denoising autoencoders can be stacked to initialize deep architectures. The algorithm can be motivated from a manifold learning and information theoretic perspective or from a generative model perspective. Comparative experiments clearly show the surprising advantage of corrupting the input of autoencoders on a pattern classification benchmark suite.