세미나

혁신적인 연구에 관심 있는 학부생, 대학원생 여러분의 많은 참여를 환영합니다!

📅 세미나 정보

🔸일시 : 2025년 2월 26일(수) 10:30

🔸장소 :  110동 N104호

🔸발표자 : Minsu Kim (Postdoctoral Researcher, KAIST–Mila Prefrontal Research Center)

🔸강연주제: Controllable Generation with GFlowNets for Vision, Language, Drug Discovery, and Combinatorial Optimization.

  – This seminar will be conducted in English

📑Abstract:

Modern deep generative models can produce highly realistic outputs across various domains, including vision (e.g., Stable Diffusion 3), language (e.g., GPT-4), drug discovery (e.g., AlphaFold 3), and combinatorial optimization (e.g., DIFFUSCO). Yet, guiding these models to yield results surpassing those found in their training data—by leveraging reward functions—remains a significant challenge.

Recently, reinforcement learning (RL) approaches, such as RL with human feedback (RLHF) and verifier-based RL for reasoning (e.g, Deepseek R1), have shown promise in controlling generative models. However, many of these methods rely predominantly on on-policy RL, which tends to be less sample efficient and more prone to mode collapse, thereby limiting output diversity.

In this talk, I will introduce an alternative perspective on controllability by framing it as a Bayesian posterior inference problem, and address it using an off-policy RL framework known as generative flow networks (GFlowNets)—an approach that can promote both output diversity and sample efficiency. I will then present my recent work on advancing GFlowNets and their application to tasks in vision, language, drug discovery, and combinatorial optimization. Finally, I will discuss the current limitations and challenges of this method, along with potential future directions in both methodology and real-world applications.

⚡연사정보: https://minsuukim.github.io/