Pengfei Li

I am Pengfei Li, a fourth-year CS Ph.D. student in University of California, Riverside, under the supervision of Prof. Shaolei Ren. Recently, we also work closely with Adam Wierman on Online Convex Optimization problems. I obtained a M.S.E degree in Laboratory for Computational Sensing and Robotics (LCSR), Johns Hopkins University, under the supervision of Prof. Alan Yuille and Prof. Gregory Hager.

Prior to joining JHU, I graduated from Zhejiang University with honors, majoring in Electrical Engineering. I was also a member of Advanced Class of Engineering Education (ACEE) in Chu Kochen Honor College (CKC). In summer 2017, I took part in the International Summer Research Program in UCSD under the supervision of Prof. Atanasov as a research intern.

Email / Twitter / Github / CV / Google Scholar

• Two papers accepted by NeurIPS 2023 (one first-authored), see you all in New Orleans !🥳

• Our preprint about LLM water consumption has been covered by AP (Associated Press), check it out here.

• Check out our new preprint about how to achieve evironmental equity in large AI models (e.g. LLM)

My research interests mainly focus on Nonlinear Optimization, Machine Learning and Graph Theory. Representative papers are highlighted, * indicates equal contributions.

We study the most general form of Smoothed Online Convex Optimization, a.k.a. SOCO, including multi-step nonlinear switching costs and feedback delay. We propose a novel machine learning (ML) augmented online algorithm, Robustness-Constrained Learning(RCL). Importantly, RCL is the first ML-augmented algorithm with a provable robustness guarantee in the case of multi-step switching cost and feedback delay.

This paper (ACRL) studies the problem of Anytime-Competitive Markov Decision Process (A-CMDP), which optimizes the expected reward while guaranteeing a bounded cost in each round of any episode against a policy prior. Experiments on the application of carbon intelligent computing verify the reward performance and cost constraint guarantee of ACRL.

A novel RL-based approach for edge-weighted online bipartite matching with robustness guarantees, achieving both good average-case performance and strict worst-case guarantee. This framework (LOMAR) supports training the RL policy by explicitly considering the online robustification operation.

The growing carbon footprint of large artificial intelligence (AI) models, such as GPT-3, has been undergoing public scrutiny. Unfortunately, however, the equally important and enormous water footprint of AI models has remained under the radar. We highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly sustainable AI.

We propose a novel expert-robustified learning (ERL) approach, with a novel projection operator. Our method can be easily extended to multistep memory costs, achieving both good average performance and strict robustness.

EC-L2O is the first to address the "how to learn" challenge for online convex optimization, which requires new algorithm design, closed-form differentiation and theoritical analysis.

A novel 3D crowd human pose estimation method, which proposed a faster cross-view matching based on graphical model

A technique to interpret 3D object detection results from volumetric-based networks

Estimated car pose and activity based on global information (ground plane parameter) estimation. Copy real car's behavior into a game engine, UE4.

©Pengfei Li, last updated Sep 2023.
Template from Jon Barron.