A brief introduction to deep learning for generative modeling
Visual recognition has experienced drastic transformations due to the adoption of machine learning techniques since the early 2000’s, and in particular the widespread adoption of deep Convolutional Neural Networks (CNNs) since 2012. State of the art approaches for tasks such as object detection, image retrieval, or semantic segmentation are now invariably based on CNNs. While extremely successful, supervised deep learning techniques such as CNNs have a number of limitations, including the requirement of large labelled training data sets. In this tutorial we focus on unsupervised deep learning techniques for visual data. These are of interest for a number of reasons, including (i) to learn visual representations without the need for supervised training data, (ii) to learn generative models to synthesize new images, (iii) as a tool to allow structured prediction in supervised tasks. This tutorial covers the basic principles of deep learning and gives an overview of the main paradigms in unsupervised deep learning, including generative adversarial networks, variational auto-encoders, autoregressive models, and invertible flow based models.