Training Neural Networks

Speaker: 

Zhiqin Lu

Institution: 

UC Irvine

Time: 

Friday, May 11, 2018 - 4:00pm

Location: 

MSTB 120

In this talk, we shall give a mathematical setting of the Random Backpropogation  (RBP) method in unsupervised machine learning. When there is no hidden layer in the neural network, the method degenerates to the usual least square method. When there are multiple hidden layers, we can formulate the learning procedure as a system of nonlinear ODEs. We proved the short time, long time existences as well as the convergence of the system of nonlinear ODEs when there is only one hidden layer. This is joint work with Pierre Baldi in Neural Networks 33 (2012) 136-147, and with Pierre Baldi, Peter Sadowski in Neural Networks 95 (2017) 110-133 and in Artificial Intelligence 260 (2018), 1-35.

An introduction to ABP estimate

Speaker: 

Xiangwen Zhang

Institution: 

UC Irvine

Time: 

Friday, May 18, 2018 - 4:00pm

Location: 

MSTB 120

The celebrated Alexandrov-Bakelman-Pucci Maximum Principle (often abbreviated as ABP estimate) is a pointwise estimate for solutions of elliptic equations, which was introduced in the 1960s. It was motivated by beautiful geometric ideas and has been a fundamental tool in the study of non-divergent PDEs. More recently, this PDE technique also pays back to geometry - the ABP estimate and its extensions can be used to prove some optimal classical geometric inequalities such as the Isoperimetric and Minkowski inequalites.

Random Integer Matrices and Random Finite Abelian Groups

Speaker: 

Nathan Kaplan

Institution: 

UC Irvine

Time: 

Friday, June 1, 2018 - 3:00pm to 4:00pm

Location: 

MSTB 118

How do you choose a random finite abelian group?

A d x d integer matrix M gives a linear map from Z^d to Z^d. The cokernel of M is Z^d/Im(M). If det(M) is nonzero, then the cokernel is a finite abelian group of order det(M) and rank at most d.

What do these groups ‘look like’? How often are they cyclic? What can we say about their p-Sylow subgroups? What happens if instead of looking at all matrices, we only consider symmetric ones? We will discuss distributions on finite abelian p-groups, focusing on ones that come from cokernels of families of random matrices. We will explain how these distributions are related to questions from number theory about ideal class groups, elliptic curves, and sublattices of Z^d.

Neoclassical Theory of Electromagnetic Interactions

Speaker: 

Alexander Figotin

Institution: 

UC Irvine

Time: 

Friday, January 26, 2018 - 4:00pm

Location: 

MSTB 120

The theory of electromagnetic (EM) phenomena known as electrodynamics is one of the major theories in science. At macroscopic scales the interaction of the EM field with matter is described by the classical electrodynamics based on the Maxwell-Lorentz theory. Many of electromagnetic phenomena at microscopic scales are covered by the so-called semiclassical theory that treats the matter according to the quantum mechanics, whereas the EM field is treated classically. The subject of this presentation is a recently advanced by us neoclassical electromagnetic theory that describes EM phenomena at all spatial scales –microscopic and macroscopic. This theory modifies the classical electrodynamics into a theory that applies to all spatial scales including atomic and nanoscales. The neoclassical theory is conceived as one theory for all spatial scales in which the classical and quantum aspects are naturally unified and emerge as approximations. It is a classical Lagrangian field theory, and consequently it is a local and deterministic theory. Probabilistic aspects of the theory may arise in it effectively through complex nonlinear dynamical evolution. This presentation is to provide an introduction to our theory including a concise historical review.

(Joint work with Anatoli Babin)

Pages

Subscribe to RSS - Graduate Seminar