|
I am a 4th-year Ph.D. student in Electrical Engineering at University of California, Riverside (UCR).
I conduct research at natural language processing (NLP) and deep learning (DL).
Specifically, I'm interested in uncertainty and robustness of language models, as well as efficient training or inference of various DL tasks.
Email  
|  CV  
|  Google Scholar  
|  GitHub  
|  LinkedIn
|  Twitter
|
Education
Ph.D.
|
University of California, Riverside (UCR), Electrical and Computer Engineering, U.S.
Advisor: Cong Liu
Sep 2022 - now
|
|
M.S.
|
University of California, San Diego (UCSD), Electrical and Computer Engineering, U.S.
Sep 2018 - Jun 2020
|
|
B.S.
|
Xi'an Jiaotong University (XJTU), Mechanical Engineering, China
Sep 2014 - Jun 2018
|
|
|
Publications (* denotes equal contribution)
|
RT-LM: Uncertainty-Aware Resource Management for Real-Time Inference of Language Models
Yufei Li, Zexin Li, Wei Yang, Cong Liu
RTSS 2023
pdf
In real-world cloud servers, LLMs such as ChatGPT may need to process hundreds of queries in a short period, and how to multitask efficiently affects the overall system performance. We propose RT-LM which leverages uncertainty as a latency heuristic to optimize real-time LLM-based dialogue systems.
|
|
R^3: On-device Real-Time Deep Reinforcement Learning for Autonomous Robotics
Zexin Li, Aritra Samanta, Yufei Li, Andrea Soltoggio, Hyoseung Kim, Cong Liu
RTSS 2023
pdf
We present R^3, a holistic solution for managing timing, memory, and algorithm performance in on-device real-time DRL training. R^3 improves timing performance while minimizing the risk of out-of-memory (OOM) errors by using a deadline-driven feedback loop with dynamic batch sizing, an efficient memory management, and a runtime coordinator.
|
|
PIMbot: Policy and Incentive Manipulation for Multi-Robot Reinforcement Learning in Social Dilemmas
Shahab Nikkhoo, Zexin Li, Aritra Samanta, Yufei Li, Cong Liu
IROS 2023
pdf
We presents PIMbot, an approach to manipulate the reward function in multi-robot collaboration through two distinct forms: policy and incentive manipulation. PIMBot introduces a new angle for manipulating multi-agent RL social dilemmas through a unique reward function for incentivization. By using our proposed mechanisms, a robot is able to manipulate the social dilemma environment effectively.
|
|
Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data
Yufei Li, Xiao Yu, Yanchi Liu, Haifeng Chen, Cong Liu
ACL 2023
pdf | code
We propose UnBED, an uncertainty-aware bootstrap learning framework for information extraction on distantly-supervised data. Specifically, UnBED uses self-ensembling as a regularizer to mitigate inter-model uncertainty caused by noisy labels. It further applies probability variance of dropout predictions to quantify innermodel uncertainty, and iteratively builds new reliable training subset.
|
|
White-Box Multi-Objective Adversarial Attack on Dialogue Generation
Yufei Li, Zexin Li, Yingfan Gao, Cong Liu
ACL 2023
pdf | code
We propose a white-box multi-objective method called DGSlow for attaking dialogue systems. Specifically, DGSlow balances two objectives---generation accuracy and length, via a gradient-based multiobjective optimizer and applies an adaptive searching mechanism to iteratively craft adversarial samples with only a few modifications.
|
|
SHARE: a System for Hierarchical Assistive Recipe Editing
Shuyang Li, Yufei Li, Jianmo Ni, Julian McAuley
EMNLP 2022
pdf | code
We introduce SHARE: a System for Hierarchical Assistive Recipe Editing to assist home cooks with dietary restrictions---a population under-served by existing cooking resources. Our hierarchical recipe editor makes necessary substitutions to a recipe's ingredients list and re-writes the directions to make use of the new ingredients.
|
|
GLIB: Towards Automated Test Oracle for Graphically-Rich Applications
Ke Chen*, Yufei Li*, Yingfeng Chen, Changjie Fan, Zhipeng Hu, Wei Yang
ESEC/FSE 2021
pdf | code
We propose GLIB, a CNN-based game GUI glitch detection framework empowered by code-based data augmentation. Specifically, we inject buggy code snippets to the game app source-code and record the manifestation of the bugs (i.e., UI glitches). Our generated screenshots contain real UI glitches so that the DL model can be trained on more precise datasets and potentially learn subtle imperceptible patterns.
|
|
Work Experience
NEC Laboratories America, Inc., Princeton, NJ, U.S.
Research Intern • May 2022 - Aug 2022
Annotated named entities for few-shot prompt-based field extraction from log messages
Defined hierarchical relations between log components and configured dynamic attributed graphs
Detected anomalies in log messages using a GNN-based encoder enhanced with temporal-attentive transformers
|
|
NEC Laboratories America, Inc., Princeton, NJ, U.S.
Research Intern • May 2021 - Aug 2021
Annotated name entities and relations using regular expressions in CVE corpus for distant supervision
Incorporated pre-trained GPT-2 into a sequence labeling framework for information extraction (IE)
Proposed a bootstrap training strategy for denoising distant labels and selecting high-quality instances
|
|
The University of Texas at Dallas (UTD), Dallas, TX, U.S.
Research Assistant • Aug 2020 - May 2022
Work @ Dr. Wei Yang's Lab, doing research on NLP and software engineering.
|
|
SeekTruth Scientific and Technical Corporation, Beijing, China
Research Intern • Jul 2019 - Sep 2019
Built a joint key point and pose recognition model for character identification tasks
Developed an adaptive discrimination definition mode from Caffe to TensorFlow
Designed a lightweight CNN to calibrate video frame orientations in real-time for online streaming
|
|
|
Projects
|
Content-aware Dynamic Graphs for Log Anomaly Detection [code]
NLP & Data Mining • May 2022 - Feb 2023
Configured dynamic attributed graphs by identifying log components and their hierarchical relationships
Proposed a GNN-based temporal-attentive transformer for detecting anomalous edges in dynamic graphs
|
|
Distantly-supervised Joint Entity and Relation Extraction with Noise-robust Learning [code]
NLP • Sep 2021 - Present
Incorporated a pre-trained transformer into sequence tagging scheme for distantly-supervised joint extraction
Proposed a bootstrap learning framework with a noise-robust loss to dynamically select high-quality instances
|
|
GAET: Assessing the Reusability of Pre-trained Code Embeddings [code]
NLP & SE • Sep 2020 - May 2021
Developed a cost-efficient offline framework to assess the generalizability of embeddings in code analysis tasks
Evaluated the generalizability of existing pre-trained embeddings leveraging semantic metamorphic relationships
|
|
Rethink Negative Sampling in Bayesian Personalized Ranking [code]
Recommender Systems • Nov 2019 - Jun 2020
Identified a limitation of popularity-based sampling due to non-uniform negative sampling biases
Rectified biases by creating tailored negative sampling distributions to boost Bayesian personalized ranking
|
|
Automatic Delivery Vehicle Design [code]
Algorithm • Mar 2019 - Jun 2019
Simulated a project integrating the Courier and TSP challenges for autonomous delivery vehicle design
Formulated a path planning algorithm by incorporating the A* heuristic rules with genetic evolution principles
|
|
Honors & Awards
VEX Robotics International Competitions
Excellent Award and Runner-Up at the VEX Robotics World Championship 2017, Louisville, KY, U.S.
Excellent Award and Runner-Up at the VEX Robotics Asia Open 2016, Beijing, China
First-class Award at the VEX Robotics China Open 2016, Xi'an, China
|
|
Scholarship Awards
National Encouragement Scholarship 2015-2017
|
|
|
Program Committee & Reviewer
|
|