# Applied Math Seminar

## Applied Math Seminar

## Applied Math Seminar

## Applied Math Seminar

## Applied Math Seminar

## Applied Math Seminar

**Title:**Exponential convergence rates for Batch Normalization

**Abstract:** Batch Normalization is a normalization technique that has been used in training deep Neural Networks since 2015. In spite of its empirical benefits, there exists little theoretical understanding as to why this normalization technique speeds up learning. From a classical optimization perspective, we will discuss specific problem instances in which we can prove that Batch Normalization can accelerate learning, and how this acceleration is due to the fact that Batch Normalization splits the optimization task into optimizing length and direction of parameters separately.

## Applied Math Seminar

**Title:**On Toric Ideals of some Statistical Models.

**Abstract:**We introduce hierarchical models from statistics and their associated Markov bases. These bases are often large and difficult to compute. We introduce certain toric ideals and their algebraic properties as an alternative way of thinking about these objects. One challenge is to describe hierarchical models with infinitely many generators in a finite way. Using a symmetric group action, we describe certain classes of models including progress made for the non-reducible Models. This is joint work with Uwe Nagel.

## SIAM Guest Speaker

**Title:** Efficient Methods for Enforcing Contiguity in Geographic Districting Problems

**Abstract: **Every ten years, United States Congressional Districts must be redesigned in response to a national census. While the size of practical political districting problems is typically too large for exact optimization approaches, heuristics such as local search can help stakeholders quickly identify good (but suboptimal) plans that suit their objectives. However, enforcing a district contiguity constraint during local search can require significant computation; tools that can reduce contiguity-based computations in large practical districting problems are needed. This talk introduces the geo-graph framework for modeling geographic districting as a graph partitioning problem, discusses two geo-graph contiguity algorithms, and applies these algorithms to the creation of United States Congressional Districts from census blocks in several states. The experimental results demonstrate that the geo-graph contiguity assessment algorithms reduce the average number of edges visited during contiguity assessments by at least three orders of magnitude in every problem instance when compared with simple graph search, suggesting that the geo-graph model and its associated contiguity algorithms provide a powerful constraint assessment tool to political districting stakeholders. Joint work with Douglas M. King and Edward C. Sewell

## Applied Math Seminar

## Applied Math Seminar

**Title:**A Mathematical Model for the Force and Energetics in Competitive Running

**Abstract:**Competitive running has been around for thousands of years and many people have wondered what the optimal form and strategy is for running a race. In his paper, Behncke develops a simple mathematical model that focuses on the relationships and dynamics between the forces and energetics at play in order to find an optimal strategy for racing various distances. In this talk, I will describe the biomechanics, energetics, and optimization of running in Behncke's model and present his findings. Note: you do not have to like running to come to this talk :)

## Applied Math Seminar

**Title:**Mathematical deep learning for drug discovery

**Abstract:**Designing efficient drugs for curing diseases is of essential importance for the 21

^{st}century's life science. Computer-aided drug design and discovery has obtained a significant recognition recently. However, the geometric complexity of protein-drug complexes remains a grand challenge to conventional computational methods, including machine learning algorithms. We assume that the physics of interest of protein-drug complexes lies on low-dimensional manifolds or subspaces embedded in a high-dimensional data space. We devise topological abstraction, differential geometry reduction, graph simplification, and multiscale modeling to construct low-dimensional representations of biomolecules in massive and diverse datasets. These representations are integrated with various deep learning algorithms for the predictions of protein-ligand binding affinity, drug toxicity, drug solubility, drug partition coefficient and mutation induced protein stability change, and for the discrimination of active ligands from decoys. I will briefly discuss the working principle of various techniques and their performance in D3R Grand Challenges,a worldwide competition series in computer-aided drug design and discovery (http://users.math.msu.edu/users/wei/D3R_GC3.pdf).