Byeongjin Kang

I'm an undergraduate in Sungkyunkwan University, majoring in Computer Science and Engineering. I'm interested in AI-research problems that hard to solve in real-world. Luckly, I worked as an intern at RLLAB (advised by Prof. Youngwoon Lee) and CSI LAB (advised by Prof. Yusung Kim)

Email  /  Github  /  Linkedin

profile photo

Research Interests

My research focuses on representation learning, aiming to create AI systems that are not just accurate but also practical for real-world use. The goal is to develop models that can learn from various data sources, generalize well, and operate efficiently on limited hardware.

My current research focuses on the following key approaches:
  • Understanding Real-world: My research aims to enhance AI's ability to perceive and understand the real world by integrating sensory information from various modalities. This involves developing algorithms that can process and interpret data from cameras, microphones, and other sensors, enabling AI systems to gain a more comprehensive understanding of their environment.

  • Self-Supervised Learning: Manual data labeling is a major bottleneck in AI development. I investigate self-supervised learning methods that let models learn powerful features from vast amounts of unlabeled data. This approach drastically reduces our reliance on costly datasets, making the training process far more efficient and scalable.

  • Knowledge Distillation: Deploying powerful AI models on resource-constrained devices is a major challenge. My work in knowledge distillation addresses this by transferring expertise from large, complex teacher models to smaller, more efficient student models, making high-performance AI practical for devices like robots or smartphones. Additionally, I investigate how to distill knowledge between models with different input distributions, allowing the student to learn a more robust representation and generalize better to its unique data domain.
By integrating these research directions, I aspire to develop an integrated AI framework that leverages multi-modal and efficient representations to address the complex, real-world challenges including Physical AI and other difficult domains.

Research Experinces and Projects
Complex Long horizontal task with Online Reinforcement Learning
Byeongjin Kang, CSI Lab team members
CSI LAB research internship project

This project explored online reinforcement learning for a complex long horizontal task using a main board and cables. It involved setting up a robust learning environment, efficient data collection, and training with pixel-based RLPD. While full success in online learning was not achieved, the project provided valuable insights into the challenges of real-time adaptation and laid the groundwork for future advancements in online reinforcement learning.

Visual Robustness in Imitation Learning
Byeongjin Kang, CSI Lab team members
CSI LAB research internship project

This project focuses on applying action imitation learning to robotic manipulation tasks, aiming for visual robustness. It tests variations in the number of cameras and viewpoints while leveraging various vision techniques to ensure resilience against visual disturbances such as color changes, brightness variations, and blur in different environments. Experiments demonstrated a significant improvement in success rates, with the baseline model ACT achieving 10% in the training environment and 80% in different test environments.

Efficient Long Text Summarization Using an sLLM Pipeline
Byeongjin Kang, HoJae Kim, HunTae Kim, Joonyeol Choi, Minsu Kim, Saehun Chun
project paper / code

This project develops a compact, edge-deployable LLM for summarizing video lecture transcripts efficiently. Instead of direct fine-tuning, it uses a segmentation-based approach, dividing transcripts into semantically related segments via cosine similarity. These are clustered using methods like KNN and DBSCAN, then summarized by a specialized 500M-parameter LLM. This approach reduces computational demands while maintaining high-quality summaries. Evaluation with ROUGE and BERTScore shows superior performance over baseline models.

Website template from Jon Barron.