Date of Award

6-26-2026

Date Published

July 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Garrett Katz

Keywords

Constrained Optimization;Full-capacity Learning;Loss Manifold Navigation;Multi-Objective Learning;Single-pass Learning;WiFi Sensing

Subject Categories

Computer Sciences | Physical Sciences and Mathematics

Abstract

Neural intelligence is identified with complex learning problems optimized over the high-dimensional, non-convex parameter spaces of deep neural networks. Solving such problems generally requires handling competing objectives and conflicting constraints. This is traditionally dealt with using heuristic methods that collapse such complexity into unconstrained, singular objectives. While computationally convenient, this invites a tradeoff against the precision and stability afforded by non-heuristic, geometry-aware approaches. This dissertation explores constrained multi-objective learning in various applied and theoretical contexts and highlights the feasibility of approximate methods as well as the necessity of exact methods. We propose a framework for equality-constrained deep learning via approximate exterior point traversal. By treating constraints not as soft penalties, but as hard geometric level sets, we uncover a numerical traversal method for fixed-level training loss manifolds with applications in loss regularization and loss landscape analysis. We also give a performance-preserving pruning method with performance-constrained updates in the parameter space, further demonstrating the utility of this constrained manifold navigation approach. Our work includes a co-first-authored study that contrasts this practical approach to reclaiming model efficiency and establishes theoretical limits for full-capacity single-pass learning in a linear perceptron. Our other contributions are included under a collection of additional studies that all share a constrained and multi-objective learning core. These include hybrid discrete-continuous learning in a WiFi sensing application, and multi-objective symbolic pattern mining in a state-space search context.

Access

Open Access

Download

Share

COinS