Jilin University · Computer Science and Technology · Class of 2023
Lee Ho
University
Jilin University
College
Computer Science and Technology
Major
Computer Science and Technology
Cohort
Class of 2023
Hometown
Hulunbuir, Inner Mongolia
Phone (WeChat)
188 0499 9997
Email
leehouc@163.com
Lee Ho · 2026
Education
Solid academic performance and strong English proficiency
Jilin University
Undergraduate, Class of 2023
College of Computer Science and Technology · Major in Computer Science and Technology
Ranked 11 / 345 in dynamic comprehensive major ranking, consistently placing near the top of the major.
Machine Learning 98
Data Structures 97.5
Microcomputer Systems 97
Compiler Principles 96.9
Microcomputer Interface Lab 96.8
Digital Logic Circuits 96.6
Computer Networks 96.5
Operating Systems 95.8
Intro to CS 95.8
Hardware Design Lab I 95.5
Open Innovation Lab I 95
English and Overall Competence
CET-4
612
CET-6
605
Strong in English literature reading and communication, supporting research investigation, paper writing, and academic exchange.
Project
Outstanding project practice and strong engineering capability
Mar 2025 - Aug 2025
Primary Project Lead
Connect6 Game System Based on an Attention-Enhanced Policy-Value Residual Network (Attention-ResNet)
As the primary project lead, I led the development of a Connect6 game system, first advancing through campus selection with Monte Carlo Tree Search (MCTS), and then upgrading the architecture for the national finals into a game-decision system built on a deep residual network (ResNet).
Project Homepage
https://github.com/Leehouc/AIDA-AFTERLIFE
- The input uses multi-channel board representations, while the backbone employs a 34×128 deep residual network to extract spatial features and combines spatial and coordinate attention for fine-grained modeling of critical regions
- The output includes both a policy head and a value head for predicting the next-move distribution and the winning trend of the current board; for the special double-stone rule in Connect6, I introduced legality constraints and key-move recognition modules to significantly reduce search explosion under complex rules
- The final model achieved 73.42% policy top-1, 91.49% policy top-5, and a value MSE of 0.221 on the validation set, ranked first with an 86.2% win rate in the Jilin University Connect6 ranking tournament, and won National First Prize in the 2025 Chinese Collegiate Computer Game Competition
Research
Strong academic research capability and innovation potential
Emotion Recognition Research Based on Reinforcement Learning and Chain-of-Thought Optimization
To address hallucinated reasoning caused by incomplete cross-modal information when the strong baseline VideoAuto-R1 faces missing audio at test time in multimodal video emotion recognition, I proposed a GRPO-based reinforcement-learning optimization framework.
- Designed an episode-level reasoning-chain structure to explicitly segment the thinking process, and trained acoustic and semantic scoring models to build a semantic reward mechanism that distinguishes and rewards or penalizes visual speculation versus unsupported hallucination
- Introduced perturbations and computed hidden-state stability rewards, integrating them into a token-level comprehensive reward allocation strategy
- Reduced over-reliance on the final target during training and provided positive feedback when the final answer was wrong but the reasoning logic was correct, effectively improving model generalization
Zero-shot Learning for Multi-label Software Vulnerability Detection
As the primary project lead, I built a unified representation framework that combines code semantic modeling with label semantic alignment to address class imbalance in multi-label vulnerability detection.
- To improve discrimination on long-tail classes in multi-label detection, I introduced ranking loss together with bias correction and per-class threshold optimization
- On seen classes, compared with the LineVul baseline, the method achieved relative improvements of 2.1% in Micro-F1 and 4.6% in Macro-F1, while reducing Hamming Loss by 7.2%
- For unseen vulnerability types, I introduced a label semantic alignment mechanism and an inference-stage calibration strategy; compared with the baseline bottleneck on unseen classes (Macro-F1 of only 0.03), the model reached a Macro-F1 of 0.32 and an accuracy of 0.42
Awards
Awards and Recognition
National First Prize, Connect6 Group, Chinese Collegiate Computer Game Competition
National Second Prize, Programming Skills Competition, RAICOM Robotics Developer Competition
National Scholarship
Outstanding Student of Jilin University
First-Class Scholarship
Dongrong Scholarship of Jilin University
Contact