Computer Science and Discrete Mathematics (CSDM)

Theoretical Computer Science and Discrete Mathematics

Operator Scaling via Geodesically Convex Optimization, Invariant Theory and Polynomial Identity Testing

Yuanzhi Li
Princeton University
March 19, 2018

We propose a new second-order method for geodesically convex optimization on the natural hyperbolic metric over positive definite matrices. We apply it to solve the operator scaling problem in time polynomial in the input size and logarithmic in the error. This is an exponential improvement over previous algorithms which were analyzed in the usual Euclidean, “commutative” metric (for which the above problem is not convex).

Abstract Convexity, Weak Epsilon-Nets, and Radon Number

Shay Moran
University of California, San Diego; Member, School of Mathematics
March 13, 2018

Let F be a family of subsets over a domain X that is closed under taking intersections. Such structures are abundant in various fields of mathematics such as topology, algebra, analysis, and more. In this talk we will view these objects through the lens of convexity.
We will focus on an abstraction of the notion of weak epsilon nets:
given a distribution on the domain X and epsilon>0,
a weak epsilon net for F is a set of points that intersects any set in F with measure at least epsilon.

On the Communication Complexity of Classification Problems

Roi Livni
Princeton University
February 27, 2018

We will discuss a model of distributed learning in the spirit of Yao's communication complexity model. We consider a two-party setting, where each of the players gets a list of labelled examples and they communicate in order to jointly perform a learning task. For example, consider the following problem of Convex Set Disjointness: In this instance Alice and Bob each receive a set of examples in Euclidean space and they need to decide if there exists a hyper-plane that separate the sets.

A Tight Bound for Hypergraph Regularity

Guy Moshkovitz
Harvard University
February 26, 2018

The hypergraph regularity lemma — the extension of Szemeredi's graph regularity lemma to the setting of k-graphs — is one of the most celebrated combinatorial results obtained in the past decade. By now there are various (very different) proofs of this lemma, obtained by Gowers, Rodl et al. and Tao. Unfortunately, what all these proofs have in common is that they yield partitions whose order is given by the k-th Ackermann function.

Some closure results for polynomial factorization

Mrinal Kumar
Harvard University
February 20, 2018

In a sequence of extremely fundamental results in the 80's, Kaltofen showed that any factor of n-variate polynomial with degree and arithmetic circuit size poly(n) has an arithmetic circuit of size poly(n). In other words, the complexity class VP is closed under taking factors.
A very basic question in this context is to understand if other natural classes of multivariate polynomials, for instance, arithmetic formulas, algebraic branching programs, bounded depth arithmetic circuits or the class VNP, are closed under taking factors.

Nonlinear dimensionality reduction for faster kernel methods in machine learning.

Christopher Musco
Massachusetts Institute of Technology
February 12, 2018

The Random Fourier Features (RFF) method (Rahimi, Recht, NIPS 2007) is one of the most practically successful techniques for accelerating computationally expensive nonlinear kernel learning methods. By quickly computing a low-rank approximation for any shift-invariant kernel matrix, RFF can serve as a preprocessing step to generically accelerate algorithms for kernel ridge regression, kernel clustering, kernel SVMs, and other benchmark data analysis tools.

Outlier-Robust Estimation via Sum-of-Squares

Pravesh Kothari
February 6, 2018

We develop efficient algorithms for estimating low-degree moments of unknown distributions in the presence of adversarial outliers. The guarantees of our algorithms improve in many cases significantly over the best previous ones, obtained in recent works. We also show that the guarantees of our algorithms match information-theoretic lower-bounds for the class of distributions we consider. These better guarantees allow us to give improved algorithms for independent component analysis and learning mixtures of Gaussians in the presence of outliers.