A longstanding challenge in machine learning and computational neuroscience is to characterise the statistics of natural images. Traditionally this has been addressed by attempting to learn the statistics of small image patches. In recent years, deep generative models such as GANs have shown the ability to generate remarkably detailed full size natural images from a random input. Yet the extent to which such models capture the statistics of full images is poorly understood.

In this work we present a simple method to evaluate generative models of full images. Applying our method to GANs show that they fail to capture very basic properties of the distribution. We also show how it is possible to learn Gaussian Mixture Models (GMMs) that can generate rich and detailed full images despite the high dimensionality and that in many respects, the GMMs capture the distribution better than GANs.

Bio:

Yair Weiss is a Professor at the School for Computer Science and Engineering at the Hebrew University. For the last four years he is also serving as Dean of the School. He received his PhD from the Massachusetts Institute of Technology in 1998, did postdoctoral work at UC Berkeley and joined the faculty of the Hebrew University in 2001. His research interests include Human and Machine Vision, Neural Computation and Machine Learning. He served as the Program Chair of the Neural Information Processing Systems (NIPS 2004) conference and is currently one of the Program Chairs of the European Conference on Computer Vision (ECCV 2018). Together with his students and colleagues he has received Best Paper Awards at ECCV, CVPR, NIPS and UAI. He has also received the Michael Bruno Memorial Award and is a Senior Fellow of the Canadian Institute for Advanced Research.