Georgia Institute of Technology College of Sciences

I will discuss a general framework for studying what can be said about the rank of a matrix A over a field K if we only know certain crude features of A. For example, what can we say about rank(A) if we only know which entries are zero and which are nonzero? Or if K = R, what if we only know the signs of the entries of A? Or K is a normed field and we only know the absolute values? Or K=C and we only know the arguments? There are many partial answers to questions like this scattered throughout the literature, and I will explain how at least some of these results can be unified through a theory of ranks of matrices over hyperfields. This is work in progress with Tianyi Zhang.

In close collaboration with experimentalists and clinicians, mathematical models that are parameterized with experimental and clinical data can help estimate patient-specific disease dynamics and treatment success. This positions us at the forefront of the advent of ‘virtual trials’ that predict personalized optimized treatment protocols. I will discuss a couple of different projects to demonstrate how to integrate calculus into clinical decision making. I will present a variety of mathematical model that can be calibrated from early treatment response dynamics to forecast responses to subsequent treatment. This may help us to identify patient candidates for treatment escalation when needed, and treatment de-escalation without jeopardizing outcomes.

Recording link: https://bluejeans.com/s/dcDrDQuxm2W

The $n$-th ordered configuration space of a graph parametrizes ways of placing $n$ distinct and labelled particles on that graph. The homology of the one-point compactification of such configuration space is equipped with commuting actions of a symmetric group and the outer automorphism group of a free group. We give a cellular decomposition of these configuration spaces on which the actions are realized cellularly and thus construct an efficient free resolution for their homology representations. As our main application, we obtain computer calculations of the top weight rational cohomology of the moduli spaces $\mathcal{M}_{2,n}$, equivalently the rational homology of the tropical moduli spaces $\Delta_{2,n}$, as a representation of $S_n$ acting by permuting point labels for all $n\leq 10$. This is joint work with Christin Bibby, Melody Chan, and Nir Gadish. Our paper can be found on arXiv with ID 2109.03302.

Large-scale machine learning models are trained by parallel (stochastic) gradient descent algorithms on distributed systems. The communications for gradient aggregation and model synchronization become the major obstacles for efficient learning as the number of nodes and the model's dimension scale up. In this talk, I will introduce several ways to compress the transferred data and reduce the overall communication such that the obstacles can be immensely mitigated. More specifically, I will introduce methods to reduce or eliminate the compression error without additional communication.