• Keywords: physics-aware generative models, rare event analysis, learning theory

• Skills: stochastic processes, general statistical mechanics, stochastic thermodynamics, large deviation theory 

• e-mail address: yoh.phys + $at$ + gmail + $dot$ + com

• Google Scholar: link / LinkedIn: link

I am an early-career postdoc studying the intersection of statistical physics and artificial intelligence. I am currently enrolled in InnoCORE fellowship program (Hyperscale language model innovation research group) hosted by Prof. Jaejun Yoo (Lab homepage: LAIT), Ulsan National Institute of Science and Technology (UNIST), South Korea.

I received my Ph.D. in nonequilibrium statistical physics at Seoul National University (SNU), South Korea (advisor: Prof. Yongjoo Baek, Lab homepage: CANONEQal ensemble). Mainly focusing on the collective phenomena and unconventional mechanical properties of active matter and how are they sustained and modulated by various nonequilibrium flows, my Ph.D. thesis is titled "Statistics of matter, energy and information flows in active matter".  My main skills are general statistical mechanics, stochastic thermodynamics, modeling  and calculations of stochastic processes, and rare event analysis through the lens of large deviation theory.

In my postdoc, based on my understandings and skills on the geometry of solution space (mostly from the statistical physics for disordered systems), I am seeking how are the robust and well-generalizing AI models are distributed and connected each other. This line of research can be linked to recent topics on deep learning such as mode connectivity, federated learning and weight-space learning. Moreover, it can be related to broader theoretical topics on computer science including algorithmic hardness. This theoretical approach is also capable of evaluating the quality of high-dimensional neural representations and their compositionality based on the toolkit of phase transitions.

My another line of research is rare event analysis for large models. Using the tool of large deviation theory that I have been trained during my Ph.D., I can systematically amplify and investigate rare but plausible samples from a given probability distribution, ranging from nonequilibrium physical systems and practical generative pre-trained models.

I am also currently working on physics-aware generative models, aiming to build AI models which is aligned or constrained to satisfy the laws of physics.  Especially, integrating the physical laws to large-scale AI may require detailed considerations on various levels and modalities (e.g., particles, pixels, etc.).

Furthermore, as a potential topic of research, I am interested in theoretically tackling the current big challenges of AI, such as continual learning, unlearning and unconventional computing.