Georgia Institute of Technology College of Sciences

We present results on the number of linear regions of the functions that can be represented by artificial feedforward neural networks with maxout units. A rank-k maxout unit is a function computing the maximum of k linear functions. For networks with a single layer of maxout units, the linear regions correspond to the regions of an arrangement of tropical hypersurfaces and to the (upper) vertices of a Minkowski sum of polytopes. This is joint work with Guido Montufar and Leon Zhang.

Meeting link: https://bluejeans.com/912860268/9947

The Navier-Stokes and Euler equations are the fundamental models for describing viscous and inviscid fluids, respectively. Based on ideas which date back to Kolmogorov and Onsager, solutions to these equations are expected to dissipate energy, which in turn suggests that such solutions are somewhat rough and thus only weak solutions. At these low regularity levels, however, one may construct wild weak solutions using convex integration methods. In this talk, I will discuss the motivation and methodology behind joint work with Tristan Buckmaster, Nader Masmoudi, and Vlad Vicol in which we construct wild solutions to the Euler equations which deviate from the predictions of Kolmogorov's classical K41 phenomenological theory of turbulence.

Random graphs with latent geometric structure, where the edges are generated depending on some hidden random vectors, find broad applications in the real world, including social networks, wireless communications, and biological networks. As a first step to understand these models, the question of when they are different from random graphs with independent edges, i.e., Erd\H{o}s--R\'enyi graphs, has been studied recently. It was shown that geometry in these graphs is lost when the dimension of the latent space becomes large. In this talk, we focus on the case when there exist different notions of noise in the geometric graphs, and we show that there is a trade-off between dimensionality and noise in detecting geometry in the random graphs.

Designing mechanisms to maximize revenue is a fundamental problem in mathematical economics and has various applications like online ad auctions and spectrum auctions. Unfortunately, optimal auctions for selling multiple items can be unreasonably complex and computationally intractable. In this talk, we consider a revenue-maximizing seller with n items facing a single unit-demand buyer. Our work shows that simple mechanisms can achieve almost optimal revenue. We approached the tradeoffs of simplicity formally through the lens of computation and menu size. Our main result provides a mechanism that gets a (1 − ε)-approximation to the optimal revenue in time quasi-polynomial in n and has quasi polynomial (symmetric) menu complexity. 

Joint work with Pravesh Kothari, Ariel Schvartzman, Sahil Singla, and Matt Weinberg.