Event
Ph.D. Dissertation Defense: Chenghao Deng
Monday, May 11, 2026
12:00 p.m.
IRB 5105
ANNOUNCEMENT: Ph.D. Dissertati
Name: Chenghao Deng
Committee:
Committee:
Professor Furong Huang(Chair)
Professor Kaiqing Zhang
Professor Ang Li
Professor Tom Goldstein
Professor Haizhao Yang (Dean's Representative)
Date/Time: Monday, May 11, 2026, 12:00-2:00 pm
Location: IRB 5105
Title: Building Machine Learning Systems Aligned with Human Values
Abstract: As machine learning (ML) systems assume an increasingly important role in modern society, it is imperative that they be developed in ways that reflect human values. In particular, aligning ML systems with human values entails ensuring that they make fair decisions, maintain robust performance under uncertainty and adversarial conditions, exhibit behavior that more closely reflects human cognition, and are deployed in ways that are socially responsible and affordable. These values should be incorporated throughout the entire ML pipeline, from data preparation and model training to test-time evaluation and system deployment.
Motivated by these goals, we investigate how to align ML systems with human values across the full development lifecycle. First, we study temporal discrimination in sequential decision-making and introduce ELBERT, a principled measure of long-term group fairness together with an effective bias mitigation method. We then improve robustness in reinforcement learning by proposing PROTECTED, a framework that combines non-dominated policy discovery during training with online adaptation at test time. Next, we examine the performance and generalization of large language models (LLMs) across different difficulty levels through Easy2Hard-Bench, a benchmark collection with standardized difficulty annotations derived from extensive human and model evaluations. Finally, we propose FlowBank, a portfolio-based framework for LLM multi-agent workflow optimization that constructs a compact set of complementary workflows and adaptively selects the most suitable workflow for each query under a performance--cost trade-off in large-scale deployment.
