Thursday, July 6, 2023, 9:30am

Strong robustness and privacy properties are emerging as essential requirements, in addition to producing optimal solutions, in many practical applications of machine learning and optimization. Consequently, a plethora of techniques exist in the literature to address these requirements. Most of these techniques are developed for either specific problems like linear regression, logistic regression, support vector machines, etc or specific classes of problems like strongly convex loss functions, generalized linear models, etc. Moreover, the two questions of robustness and privacy are usually handled using separate techniques. This makes it very challenging for a practitioner to find the correct set of techniques which are backed by theoretical guarantees when faced with novel practical circumstances, especially ones involving DNNs that specify loss functions that have no closed form description beyond being reasonably smooth.

In this thesis, we propose geometric techniques which work only under the assumption of smoothness on loss functions. This naturally covers a large class of ML problems. For this general setting we propose a single novel random walk and study its convergence properties. Moreover, we propose strong notions of robustness and privacy for this general setting and establish that our random walk satisfies these properties.

Thesis Committee:
Pradeep Ravikumar (Chair)
Virginia Smith
Ramarathnam Venkatesan (Microsoft Research)
Stephen D. Miller (Rutgers University)

Additional Information

Zoom Participation. See announcement

Event Type: Thesis Proposals
Room Number: Virtual Presentation - ET
Building: Remote Access - Zoom
Speaker's Name: KARTIK GUPTA
Speaker's Professional Title: Ph.D. Student, Machine Learning Department, Carnegie Mellon University
Talk Title: A geometrical approach to optimization in Machine Learning for robustness and privacy
For More Information: stidle@andrew.cmu.edu
Affiliations: Machine Learning Department (MLD)
Organization(s): SCS