Title: Uncertainty-Aware and Data-Efficient Fine-Tuning and Application of Foundation Models

Date: April 18th, 2025

Time: 1.00 PM — 2.30 PM, EDT

Location: CODA C1103 Lindberg; also on Zoom: https://gatech.zoom.us/j/94101453773

Yinghao Li

Machine Learning PhD Student

School of Electrical and Computer Engineering
Georgia Institute of Technology

Committee

1. Dr. Chao Zhang (CSE, Advisor)

2. Dr. Rampi Ramprasad (MSE, Co-Advisor)

3. Dr. Tuo Zhao (ISYE)

4. Dr. Srijan Kumar (CSE & Lighthouz AI)

5. Dr. Victor Fung (CSE)

6. Dr. Ali Torkamani (Amazon Web Services)

Abstract

Pre‑trained foundation models power modern natural language processing and scientific workflows, yet their deployment is hindered by training-inference distribution shifts, scarce in‑domain labels, model hallucination, and poorly calibrated model confidence. This thesis tackles these obstacles along two fronts: 1) reliable uncertainty quantification (UQ) and 2) data‑efficient model learning. For reliability, we establish MUBen, a best‑practice UQ benchmark for molecular property prediction, and propose UQAC, which backtracks attention chains to approximate the intractable marginalization over the reasoning space, yielding better calibrated answer probabilities from large language models (LLMs). To boost data efficiency, we introduce CHMM and its sparse variant for weakly supervised named entity recognition, and devise G&O, a zero‑shot information extraction framework that harnesses LLM reasoning. We further present ELREA, a fine‑tuning strategy that clusters input instructions by gradient direction for training task‑specific LoRA experts and inference-time expert ensembling, improving model generalization capability without relying on additional data points. Together, these contributions enhance the trustworthiness, robustness, and adaptability of foundation models in high‑stakes real‑world settings.