PhD Thesis Defense

Monday May 18, 2020 10:00 AM

Universality Laws and Performance Analysis of the Generalized Linear Models

Speaker: Ehsan Abbasi, Electrical Engineering, California Institute of Technology
Location:

In the past couple of decades, non-smooth convex optimization has emerged as

a powerful tool for the recovery of structured signals (sparse, low rank, etc.)

from noisy linear or non-linear measurements in a variety of applications in

genomics, signal processing, wireless communications, machine learning, etc..

Taking advantage of the particular structure of the unknown signal of interest

is critical since in most of these applications, the dimension p of the signal to

be estimated is comparable, or even larger than the number of observations n.

With the advent of Compressive Sensing there has been a very large number of

theoretical results that study the estimation performance of non-smooth convex

optimization in such a high-dimensional setting.

A popular approach for estimating an unknown signal 0 2 Rp in a generalized

linear model, with observations y = g(X 0) 2 Rn, is via solving the estima-

tor ^ = arg min L(y;X ) + f( ). Here, L(; ) is a loss function which is

convex with respect to its second argument, and f() is a regulaizer that en-

forces the structure of the unknown 0. We rst analyze the generalization error

performance of this estimator, for the case where the entries of X are drawn

independently from real standard Gaussian distribution. The precise nature of

our analysis permits an accurate performance comparison between dierent in-

stances of these estimators, and allows to optimally tune the hyperparameters

based on the model parameters. We apply our result to some of the most popu-

lar cases of generalized linear models, such as M-estimators in linear regression,

logistic regression and generalized margin maximizers in binary classi cation

problems, and Poisson regression in count data models. The key ingredient of

our proof is the Convex Gaussian Min-max Theorem (CGMT), which is a tight

version of the Gaussian comparison inequality proved by Gordon in 1988. Un-

fortunately, having real iid entries in the features matrix X is crucial in this

theorem, and it cannot be naturally extended to other cases.

But for some special cases, we prove some universality properties and indirectly

extend these results to more general designs of the features matrix X, where

the entries are not necessarily real, independent, or identically distributed. This

extension, enables us to analyze problems that CGMT was incapable of, such as

models with quadratic measurements, phase-lift in phase retrieval, and data re-

covery in massive MIMO, and help us settle a few long standing open problems

in these areas.

Series Thesis Seminar

Contact: Tanya Owen tanya@caltech.edu