Quentin Berthet
Research Scientist, Google DeepMind, Paris France
15th June 2023, 04:00pm - 05:00pm (GST)
| Title: | Perturbed Optimizers for Machine Learning |
| Abstract: | Machine learning pipelines often rely on optimizers procedures to make discrete decisions (e.g., sorting, picking closest neighbors, or shortest paths). Although these discrete decisions are easily computed in a forward manner, they break the back-propagation of computational graphs. In order to expand the scope of learning problems that can be solved in an end-to-end fashion, we propose a systematic method to transform optimizers into operations that are differentiable and never locally constant. Our approach relies on stochastically perturbed optimizers, and can be used readily within existing solvers. Their derivatives can be evaluated efficiently, and smoothness tuned via the chosen noise amplitude. We also show how this framework can be connected to a family of losses developed in structured prediction, and give theoretical guarantees for their use in learning tasks. We demonstrate experimentally the performance of our approach on various tasks, including recent applications on protein sequences. |
| Bio: | Quentin Berthet is a research scientist at Google DeepMind, in Paris. His main research interest is in using tools from statistics and optimization to improve modern machine learning methods, with a recent interest in making unconventional operations end-to-end differentiable. Before that, he was a professor in the Department of Mathematics and Mathematical Statistics, at the University of Cambridge, and a postdoctoral fellow at the California Institute of Technology. He obtained his Ph.D. at Princeton University and is an alumni of Ecole Polytechnique. |