Sébastien Bubeck
I manage the Machine Learning Foundations group at MSR Redmond. The group spans a wide variety of topics in machine learning and theoretical computer science. My best works have been around online decision making, with a couple of solutions to longstanding problems (minimax rate for multiarmed bandits and linear bandits at COLT 2009/COLT 2012/ALT 2018, best of both worlds for multiarmed bandits at COLT 2011, bandit convex optimization at COLT 2016/STOC 2017, progress on kserver and metrical task systems at STOC 2017/SODA 2018, chasing convex bodies at STOC 2019, multiplayer multiarmed bandit at COLT 2020). I also did a couple of works in convex optimization (entropic barrier at COLT 2015, geometric view on acceleration in 2015, optimal distributed rates at ICML 2017/NIPS 2018/NeurIPS 2019/ICML 2020) and in network analysis (influence of the seed in preferential attachment graphs, and dimension estimation in random geometric graphs, work done in 2013/2014, appeared in Random Structures and Algorithms). Some other fun side projects included Langevin diffusion (NIPS 2015), entropic CLT (International Mathematics Research Notices 2016), smoothed analysis of local search (STOC 2017), and finding critical points on nonconvex functions in low dimensions (COLT 2020). I now spend a lot of time thinking about adversarial examples in ML (some works that already appeared: computational complexity of adversarial examples at ICML 2019, more efficient randomized smoothing at NeurIPS 2019, construction of compact and smooth twolayers neural networks at NeurIPS 2020). NEW: A law of robustness for neural networksWe conjecture that for most datasets, any twolayers neural network that interpolate the data must have its Lipschitz constant larger than the square root of the ratio between the number of data points and the number of neurons. This would formally prove that overparametrization is necessary for robustness. We made some progress (with Yuanzhi Li and Dheeraj Nagaraj) on this question here. Multiplayer multiarmed banditThe nonstochastic version of cooperative multiplayer multiarmed bandit (with collisions) turns out to be a surprisingly challenging problem. In this first paper we obtain an optimal algorithm for the model with announced collisions. The model where collisions are not announced remains wide open (both from upper and lower bounds perspective). In the stochastic case we proved that in fact one can get optimal regret without ANY collisions. Chasing convex bodies competitivelyWith an extremely fun team of coauthors (Yin Tat Lee, Yuanzhi Li, Mark Sellke) we finally managed to obtain a competitive algorithm for chasing convex bodies (after a couple of years of infructuous attempts), see also this youtube video. We also obtained a rather complete picture of the nested version of the problem. The latter approach was then used to obtain the optimal competitive ratio in this paper by Mark Sellke (see also this paper with very similar results). A regularization approach for kserver and metrical task systems with the multiscale entropyWith a fantastic team of coauthors (Michael Cohen, James R. Lee, Yin Tat Lee, Aleksander Madry) we improved the state of the art competitive ratio for kserver and metrical task systems by using the mirror descent algorithm. To learn more about it I recommend to first take a look at this [youtube video], then these 3 blog posts ([part 1], [part 2], [part 3]) and finally [the kserver paper] itself. The [MTS paper] is finally online too, and it is a great start to get into this line of work. Polynomialtime algorithm for bandit convex optimizationFrom July 2014 to July 2016 with various coauthors at MSR we dedicated a lot of energy to bandit convex optimization. The end product is a new efficient algorithm. To learn more about it I recommend to first take a look at this [youtube video], then these 3 blog posts ([part 1], [part 2], [part 3]), and finally [the paper] itself. Research Interests
