\nIC Faculty candidate

\n

\n

\n Probability distributions are the core object of statistical machine learn ing\, and one of the core properties we can consider is distances between them. In this talk\, we will consider using these distances for two import ant tasks\, and show how to design distances which will be useful for each task. First\, we study the problem of two-sample testing\, where we wish to determine whether two different datasets meaningfully differ\, and if s o how they differ. We then study this framework in the setting of training generative models\, such as generative adversarial networks (GANs)\, whic h learn to sample from complex distributions such as those of natural imag es.

\n

\nThe distances used are defined in terms of kernels\, but we parameterize these kernels in as deep networks for flexibility. This co mbination gives both theoretical and practical benefits over staying purel y in either framework\, and we obtain state-of-the-art results for unsuper vised image generation on CelebA and ImageNet with our novel Scaled MMD GA N.

\n

\n

\nDougal Sutherland is a postdoctor al researcher at the Gatsby Computational Neuroscience Unit\, University C ollege London\, working with Arthur Gretton. He received his PhD in 2016 f rom Carnegie Mellon University\, advised by Jeff Schneider. His research f ocuses on problems of learning about distributions from samples\, includin g training implicit generative models\, density estimation\, two-sample te sting\, and distribution regression. His work combines kernel frameworks w ith deep learning\, and aims for theoretical grounding of practical result s.

\n

\n