The 3rd KAIST-Mila Prefrontal AI Workshop

The 3rd KAIST-Mila

Prefrontal AI Workshop

Overview

Welcome to the second annual KAIST-Mila Prefrontal AI Workshop. This event highlights our continued research partnership through the Prefrontal AI Research Center, supported by the National Research Foundation of South Korea.

The workshop brings together leading researchers to explore breakthrough developments in System 2 AI, Safe AI, and AI for Science, and offers an ideal setting for knowledge sharing, networking, and building lasting research partnerships.

Connect with an international community and discover opportunities for collaboration, research exchanges, and academic visits to South Korea through meaningful discussions that shape the future of AI.

⏳When: Wednesday, Aug 19th, 2 pm - 5:45 pm (EDT)

📍Where: Coworking Space at Mila (1st floor in Building 6650 rue Saint-Urbain, H2S 3G9, Montréal)

📹Zoom: https://us06web.zoom.us/j/83941189114?pwd=tHHaN2d8cvaRl1tGJK5IIWLMPLDNrI.1

Schedules

14:00 - 14:30: Siamak Ravanbakhsh (Mila & McGill University)

Questions at the Intersection of Diffusion, RL and Sampling

Abstract: TBD

14:30 - 15:00: Seunghoon Hong (KAIST)

Flow Map Language Models: One-step Language Modeling via Continuous Denoising

Abstract: TBD

15:00 - 15:30: Dhanya Sridhar (Mila & UdeM)

Towards reliable mechanistic understanding with causal representation learning

Abstract: TBD

15:30 - 16:00: Break Time (30 min) Coffee & Snack

16:00 - 16:35: Sungsoo Ahn (KAIST)

Amortizing Molecular Simulation with Neural Networks

Abstract: Molecular simulations often spend most of their computation exploring microscopic configurations, even though the final goal is to estimate a small number of macroscopic observables. In this talk, I will discuss how AI can accelerate this process. I will first discuss the increasingly popular paradigm of neural samplers for molecular conformations, which can enable importance-sampling estimates of macroscopic observables, with a particular focus on MOFs and heterogeneous catalysts. I will then introduce an alternative paradigm, with connections to classic density functional theory, that predicts the equilibrium density field directly, which can bypass explicit configurational sampling and obtain observables from a one-shot prediction of the equilibrium state. I will conclude the talk with discussion on open problems in this field and our future research direction.

16:35 - 17:10: Sungjin Ahn (KAIST)

Towards World Theory Models

Abstract: TBD

17:10 - 17:40: Yoshua Bengio (Mila & LawZero & UdeM)

Exponential safety guarantees against harms that require coordinated deviations from AI good behavior

Abstract: TBD