Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

TL;DR

This paper introduces Deep Convolutional GANs (DCGANs), a new architecture for unsupervised learning that learns hierarchical image representations and can be applied to various image tasks.

Contribution

The paper proposes architectural constraints for CNNs to create DCGANs, demonstrating their effectiveness in unsupervised learning and feature extraction.

Findings

DCGANs learn hierarchical representations from object parts to scenes.

Features learned by DCGANs are useful for various image tasks.

DCGANs outperform previous unsupervised models in representation quality.

Abstract

In recent years, supervised learning with convolutional networks (CNNs) has seen huge adoption in computer vision applications. Comparatively, unsupervised learning with CNNs has received less attention. In this work we hope to help bridge the gap between the success of CNNs for supervised learning and unsupervised learning. We introduce a class of CNNs called deep convolutional generative adversarial networks (DCGANs), that have certain architectural constraints, and demonstrate that they are a strong candidate for unsupervised learning. Training on various image datasets, we show convincing evidence that our deep convolutional adversarial pair learns a hierarchy of representations from object parts to scenes in both the generator and discriminator. Additionally, we use the learned features for novel tasks - demonstrating their applicability as general image representations.