Event
Ph.D. Research Proposal Exam: Chenghao Deng
Friday, May 2, 2025
12:00 p.m.
IRB-4107
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Chenghao Deng
Committee:
Professor Furong Huang (Chair)
Professor Sanghamitra Dutta
Professor Kaiqing Zhang
Date/time: Friday, May 2nd, 2025 at 12:00pm - 2:00 pm
Location: IRB-4107
Title: Building Machine Learning Systems Aligned with Human Values
Abstract:
With the growing deployment of machine learning (ML) systems in society, aligning these systems with human values has become essential. In this proposal, we introduce approaches for aligning human values throughout various stages of ML system development, including data preparation, training, evaluation, and deployment.
First, we address the temporal discrimination issues in sequential decision making. We propose ELBERT, a new principled measure for long-term group fairness, coupled with an effective bias mitigation approach adopted from conventional policy optimization methods.
Next, we enhance agent robustness against diverse attack scenarios in reinforcement learning (RL). We propose PROTECTED, a framework consisting of non-dominated policy discovery during training time and online adaptation during test time.
Then, we investigate the performance and generalization capabilities of large language models (LLMs) across varying difficulty levels. We introduce Easy2Hard-Bench, a collection of benchmark datasets annotated with standardized difficulty labels derived systematically from extensive human and model performance data.
Finally, we outline future research directions aimed at developing advanced reasoning pipelines for LLMs inspired by human collaborative behavior. This includes techniques for identifying optimal reasoning graph structures and designing specialized prompts for distinct roles within reasoning workflows.
