Research Interests

Nonequilibrium systems such as active matter exhibit various anomalous phenomena such as collective pattern formation, enhanced mechanical capabilities and negative responses. The possibilities of those seemingly paradoxical behaviors are mostly explained in the language of symmetry and order, often by writing down a phenomenological model and identifying a term responsible to the desired intriguing phenomenon.
In fact, active matter is sustained by various kind of flows (or currents) to be constantly kept out of equilibrium. Beyond the phenomenological modeling, I aim to explicitly model those physical flows involved in the system, so that I can tell how the strengths of desired phenomena are constrained, improved or modulated by the presence of flows.
I study how the energy flows involved in active systems provide thermodynamic upper bounds on the performance of mechanical operations and the formation of dissipative structures. In addition, information flow as a legitimate physical quantity also provides a bound of apparent violations of the 2nd law.

Studies on large deviations of stochastic systems tells us that extremely rare events tend to occur in a surprisingly structured and organized manner. The tools developed from large deviation theory also enables us to efficiently sample those 'rare but still plausible' events, adopting the formalism of stochastic optimal control theory.
Large deviation theory is also a standard tool to describe dynamical phase transition, which is the transition that is about configuration of trajectories instead of states. This transition reveals how system gradually selects different 'modes' of operation to acheive very rare events.

Since statistical physics deals with systems consisting of numerous interacting components, the structural and dynamical aspects of modern deep learning is also a legitimate topic of statistical physics. In fact, neural network can be understood as a disordered many-body system which is optimized by a stochastic dynamics, and has been natively analyzed by the tool developed from spin-glass theory and statistical field theory through decades.
While those theories often analyze the geometric structure of typical solutions, one should sometimes focus on the atypical solutions, which are reported to be easily accessible via practical gradient-based dynamics, and usually believed to have a correspondence with flat, largely connected minima which generalizes well. I am now seeking opportunities to participate in studying this topics using the tool of high-dimensional random geometry for disordered systems.

I am also interested in theoretically tacking the important questions in modern deep learning: (1) how compositional representations emerge in contemporary large models, and (2) how can we successfully do the continual learning / unlearning without destruction of the rest of the neural network capabilities.

Page updated

Google Sites

Report abuse