This Fireside Chat features Minjoon Seo, Assistant Professor at KAIST & Chief Scientist of Twelve Labs. Beyza Ermis, Research Scientist at Cohere For AI, sits down with Minjoon to discuss “Ever-evolving Language Models”
00:15 – Introducing Minjoon
01:41 – How did you get into ML research?
05:52 – How do you decide on which topics to research?
09:37 – After obtaining your degrees in the US, what was your motivation for returning to Korea?
12:38 – Research culture in US and in Korea?
14:00 – How is world knowledge encoded in language models?
20:20 – How would you define ever-evolving?
24:47 – Could you explain what catastrophic forgetting is and why it’s such a critical issue in the training of language models?
31:07 – Is forgetting always unintended or sometimes it is intended. How can we utilize forgetting?
33:30 – How can we evaluate how well a language model retains knowledge?
36:43 – What are the main challenges in developing efficient tools that effectively combine continual learning with multi-modal learning systems?
41:05 – Can you name one paper which is very influential, even influenced your career or your growth as a researcher?
45:25 – What was the most challenging period or thing in your career?
49:42 – Given the crucial role you play in the development of early-stage researchers, what advice would you offer to your peers in these positions? How do you think a good supervisor should guide their students?
52:50 – Audience Q&A
Add comment