PhD Thesis Proposal: Minh Bui

Optional ZOOM LINK
Meeting ID: 948 7106 4902
Passcode: 764842

"Geometry, Privacy, and Efficiency in Applied Machine Learning Applications"

Abstract

This dissertation is organized around three main research themes: geometric data structures for feature representation, privacy-preserving frameworks, and computational efficiency in audio signal processing.

1. Enhancing GMRA for High-Dimensional Feature Extraction: We propose improvements to the Geometric Multi-Resolution Analysis (GMRA) framework for feature extraction from high-dimensional and non-Euclidean data, including images, audio, and graph-structured data. By refining multiscale geometric approximations, this work captures the intrinsic manifold structure of complex datasets and enables more robust representation learning than standard linear dimensionality reduction methods.
2. Privacy-Preserving Federated Learning for Automatic Speech Recognition: Building on prior studies of Differential Privacy (DP) vulnerabilities in Automatic Speech Recognition (ASR) systems, this research investigates the robustness of decentralized acoustic models. We address information leakage by extending zero-order optimization attack frameworks to first-order optimization methods, improving the efficiency and accuracy of private data reconstruction from model updates.
3. Efficient Audio Signal Processing: The final theme focuses on optimizing audio signal processing for generative models and low-latency applications. We develop computationally efficient architectures that reduce algorithmic latency for high-fidelity audio synthesis, enabling real-time deployment on resource-constrained devices.

Thesis Committee

• Peter Chin (Chair/Advisor)
• Bryce Ferguson
• Wesley Marrero
• Soroush Vosoughi

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu.