Studied mathematical logic, statistics, and machine learning at Carnegie Mellon University and has a PhD in philosophy from the University of Aberdeen. He is interested in generative models of images and audio.

Speech

Topic: Generative Adversarial Networks: Introduction and Some Recent Developments

Short Description: Generative Adversarial Networks (GANs) are an exciting class of deep learning models that can generate images and other high-dimensional objects. This talk will explain the fundamentals of GANs and how refinements like Wasserstein GANs can help to stabilize training. It will also discuss interesting variations on the basic GAN concept such as InfoGAN, which can learn human-salient features such as the digit an MNIST image represents and its width and slant without any supervision whatsoever, or CycleGAN, which can can convert between two classes of images, such as images of horses and zebras, or images with and without bokeh, without seeing any paired training examples.