Computer Science Thesis Proposal

Monday, May 23, 2022 - 2:00pm to 3:30pm


In Person and Virtual - ET Traffic21 Classroom and Zoom


MIKHAIL (MISHA) KHODAK, Ph.D. StudentComputer Science DepartmentCarnegie Mellon University

The learning of algorithms and architectures

This proposal focuses on the problem of learning the parameters that specify learning algorithms, computational methods, and neural architectures. Our focus is on methods for these areas of meta-learning and automation that carry provable guarantees, can be applied in large scale applications, and work well on many different types of applications. For the problem of meta-learning algorithms we introduce a new theoretical framework called ARUBA that yields novel provable guarantees and practical methods across multiple subfields including gradient-based meta-learning, federated learning, and algorithms with prediction. For the problem or neural architecture search we study optimization problems associated with the weight-sharing paradigm to derive a new gradient-based algorithm that takes advantage of the geometric structure of the discrete relaxation; we then propose new operation spaces for discovering truly novel architectures from data and obtain human-expert-level performance on multiple diverse tasks. Thesis Committee: Maria-Florina Balcan (Co-Chair) Ameet Talwalkar (Co-Chair) Tom Mitchell Peter Bartlett (University of California, Berkeley) Piotr Indyk (Massachusetts Institute of Technology) Alexander Smola (Amazon) In Person and Zoom Participation. See announcement. 

For More Information, Contact:


Thesis Proposal