From neural PCA to deep unsupervised learning

TL;DR

This paper introduces a deep unsupervised learning network with hierarchical autoencoders and lateral shortcut connections, enabling efficient learning of invariant features across multiple levels.

Contribution

It proposes a novel hierarchical autoencoder architecture with lateral shortcuts, combining denoising autoencoder and source separation frameworks for improved deep unsupervised learning.

Findings

Demonstrates efficient training of deep networks with multi-level cost signals.

Shows the network's ability to learn invariant features.

Provides experimental validation of speedup and feature abstraction.

Abstract

A network supporting deep unsupervised learning is presented. The network is an autoencoder with lateral shortcut connections from the encoder to decoder at each level of the hierarchy. The lateral shortcut connections allow the higher levels of the hierarchy to focus on abstract invariant features. While standard autoencoders are analogous to latent variable models with a single layer of stochastic variables, the proposed network is analogous to hierarchical latent variables models. Learning combines denoising autoencoder and denoising sources separation frameworks. Each layer of the network contributes to the cost function a term which measures the distance of the representations produced by the encoder and the decoder. Since training signals originate from all levels of the network, all layers can learn efficiently even in deep networks. The speedup offered by cost terms from higher levels of the hierarchy and the ability to learn invariant features are demonstrated in experiments.