Seminars and Colloquia by Series

Applied differential geometry and harmonic analysis in deep learning regularization

Series
Applied and Computational Mathematics Seminar
Time
Monday, September 23, 2019 - 13:50 for 1 hour (actually 50 minutes)
Location
Skiles 005
Speaker
Wei ZhuDuke University

Deep neural networks (DNNs) have revolutionized machine learning by gradually replacing the traditional model-based algorithms with data-driven methods. While DNNs have proved very successful when large training sets are available, they typically have two shortcomings: First, when the training data are scarce, DNNs tend to suffer from overfitting. Second, the generalization ability of overparameterized DNNs still remains a mystery. In this talk, I will discuss two recent works to “inject” the “modeling” flavor back into deep learning to improve the generalization performance and interpretability of the DNN model. This is accomplished by DNN regularization through applied differential geometry and harmonic analysis. In the first part of the talk, I will explain how to improve the regularity of the DNN representation by enforcing a low-dimensionality constraint on the data-feature concatenation manifold. In the second part, I will discuss how to impose scale-equivariance in network representation by conducting joint convolutions across the space and the scaling group. The stability of the equivariant representation to nuisance input deformation is also proved under mild assumptions on the Fourier-Bessel norm of filter expansion coefficients.

The Jacobian Conjecture

Series
Student Algebraic Geometry Seminar
Time
Monday, September 23, 2019 - 13:15 for 1 hour (actually 50 minutes)
Location
Skiles 254
Speaker
Stephen McKeanGeorgia Tech

The Jacobian Conjecture is a famous open problem in commutative algebra and algebraic geometry. Suppose you have a polynomial function $f:\mathbb{C}^n\to\mathbb{C}^n$. The Jacobian Conjecture asserts that if the Jacobian of $f$ is a non-zero constant, then $f$ has a polynomial inverse. Because the conjecture is so easy to state, there have been many claimed proofs that turned out to be false. We will discuss some of these incorrect proofs, as well as several correct theorems relating to the Jacobian Conjecture.

An Introduction to Braids and Complex Polynomials

Series
Geometry Topology Seminar Pre-talk
Time
Monday, September 23, 2019 - 12:45 for 1 hour (actually 50 minutes)
Location
Skiles 006
Speaker
Michael DoughertyColby College

In this informal chat, I will introduce the braid group and several equivalent topological perspectives from which to view it. In particular, we will discuss the role that complex polynomials play in this setting, along with a few classical results.

Bounds on Ramsey Games via Alterations

Series
ACO Student Seminar
Time
Friday, September 20, 2019 - 13:05 for 1 hour (actually 50 minutes)
Location
Skiles 005
Speaker
He GuoMath, Georgia Tech

In this talk we introduce a refined alteration approach for constructing $H$-free graphs: we show that removing all edges in $H$-copies of the binomial random graph does not significantly change the independence number (for suitable edge-probabilities); previous alteration approaches of Erdös and Krivelevich remove only a subset of these edges. We present two applications to online graph Ramsey games of recent interest, deriving new bounds for Ramsey, Paper, Scissors games and online Ramsey numbers (each time extending recent results of Fox–He–Wigderson and Conlon–Fox–Grinshpun–He).
Based on joint work with Lutz Warnke.

Deep Generative Models in the Diffusion Limit

Series
Stochastics Seminar
Time
Thursday, September 19, 2019 - 15:05 for 1 hour (actually 50 minutes)
Location
Skiles 006
Speaker
Maxim RaginskyECE Department, University of Illinois at Urbana-Champaign

In deep generative models, the latent variable is generated by a time-inhomogeneous Markov chain, where at each time step we pass the current state through a parametric nonlinear map, such as a feedforward neural net, and add a small independent Gaussian perturbation. In this talk, based on joint work with Belinda Tzen, I will discuss the diffusion limit of such models, where we increase the number of layers while sending the step size and the noise variance to zero. The resulting object is described by a stochastic differential equation in the sense of Ito. I will first show that sampling in such generative models can be phrased as a stochastic control problem (revisiting the classic results of Föllmer and Dai Pra) and then build on this formulation to quantify the expressive power of these models. Specifically, I will prove that one can efficiently sample from a wide class of terminal target distributions by choosing the drift of the latent diffusion from the class of multilayer feedforward neural nets, with the accuracy of sampling measured by the Kullback-Leibler divergence to the target distribution.

John’s ellipsoid is not good for approximation

Series
High Dimensional Seminar
Time
Wednesday, September 18, 2019 - 15:00 for 1 hour (actually 50 minutes)
Location
Skiles 005
Speaker
Han HuangGeorgia Tech

We study the subject of approximation of convex bodies by polytopes in high dimension.  

For a convex set K in R^n, we say that K can be approximated by a polytope of m facets by a distance R>1 if there exists a polytope of P m facets such that K contains P and RP contains K. 

When K is symmetric, the maximal volume ellipsoid of K is used heavily on how to construct such polytope of poly(n) facets to approximate K. In this talk, we will discuss why the situation is entirely different for non-symmetric convex bodies.

Surface bundles in topology, algebraic geometry, and group theory

Series
Geometry Topology Student Seminar
Time
Wednesday, September 18, 2019 - 14:00 for 1 hour (actually 50 minutes)
Location
Skiles 006
Speaker
Justin LanierGeorgia Tech

I will give an introduction to surface bundles and will discuss several places where they arise naturally. A surface bundle is a fiber bundle where the fiber is a surface. A first example is the mapping torus construction for 3-manifolds, which is a surface bundle over the circle. Topics will include a construction of 4-manifolds as well as section problems related to surface bundles. The talk will be based on a forthcoming Notices survey article by Salter and Tshishiku.

A complex analytic approach to mixed spectral problems

Series
Analysis Seminar
Time
Wednesday, September 18, 2019 - 13:55 for 1 hour (actually 50 minutes)
Location
Speaker
Burak HatinoğluTexas A&M

This talk is about an application of complex function theory to inverse spectral problems for differential operators. We consider the Schroedinger operator on a finite interval with an L^1-potential. Borg's two spectra theorem says that the potential can be uniquely recovered from two spectra. By another classical result of Marchenko, the potential can be uniquely recovered from the spectral measure or Weyl m-function. After a brief review of inverse spectral theory of one dimensional regular Schroedinger operators, we will discuss complex analytic methods for the following problem: Can one spectrum together with subsets of another spectrum and norming constants recover the potential?

Species network inference under the coalescent model

Series
Mathematical Biology Seminar
Time
Wednesday, September 18, 2019 - 11:00 for 1 hour (actually 50 minutes)
Location
Skiles 006
Speaker
Hector BanosGeorgia Tech

When hybridization plays a role in evolution, networks are necessary to describe species-level relationships. In this talk, we show that most topological features of a level-1 species network (networks with no interlocking cycles) are identifiable from gene tree topologies under the network multispecies coalescent model (NMSC). We also present the theory behind NANUQ, a new practical method for the inference of level-1 networks under the NMSC.

Pages