In recent years, supervised learning with convolutional neural networks (CNNs) has seen huge adoption in computer vision applications. Comparatively, unsupervised learning with CNNs has received less attention. In this work, the authors attempt to bridge the gap between supervised learning and unsupervised learning using CNNs.
This paper introduces the concept of DCGANs, which are neural networks that are trained to distinguish interesting structures and meaning from raw image data by competing with one another. In particular, the networks that are presented appear to be a strong candidate set for unsupervised learning of features for use in realistic image generation.>>We’ll explore convincing evidence that the DCGAN pair used here learns a hierarchical representation of the features held within images ranging from objects to entire scenes. Additionally, we’ll explore the use of these learned features for novel tasks - demonstrating their applicability as generic image representations.
Matt Hardwick is a software engineer with 10+ years of experience creating software systems in various industries, including defense simulations and web-based applications. He currently leads an engineering team at Bellhops using technology to improve the way that people move from place to place. His recent software interests are Machine Learning and Elixir.
Paper: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks (https://arxiv.org/pdf/1511.06434v2.pdf)