Interpreting Machine Learning From the Perspective of Nonequilibrium Systems
summary
Presents a concrete method that turns rare-event sampling into an importance-sampling problem over trajectories by learning an optimal drift using policy gradient style training. Uses Girsanov reweighting and a KL divergence objective (decomposed into entropy production and activity) to recover the time-reversed relaxation known as the Brownian bridge. Parameterizes the drift with neural nets, uses value baselines to cut gradient variance, and demonstrates convergence on molecular-dynamics examples (Brownian/active colloids and protein folding where picosecond motions lead to hours between events) and extensibility to climate-style systems.
AI-written summary · about
chapters (11)
- 0:00Intro
- 2:23Rare but important events in chemistry
- 4:35Difficulties, simplifications and standard approaches
- 6:50The task: generate rare trajectories typically
- 10:27Reinforcement learning based approach
- 15:48Translation into physical chemistry quantities
- 18:36Value (and lower variance) in spanning disciplines
- 20:46Doob's h-theorem and optimal stochastic control
- 23:29Connections to nonequilbrium statistical mechanics
- 26:11Active Brownian motion in a potential
- 27:51Conformational changes in active Brownian solvents
related talks
01h 14mSampling Using Diffusion Processes, from Langevin to Schrödinger
01h 10mBayesian Optimization and Machine Learning for Accelerating Experiments in the Physical Sciences
01h 03mSampling Crash Course
09mPrimal--Dual Optimization and Application to Sampling
01h 11mHidden Physics Models: Machine Learning of Nonlinear Partial Differential Equations
01h 07mConsensus-Based Interacting Particle Systems and Mean-field PDEs for Optimization and Sampling
57mDarwinian Machine Learning: Principles of Machine Learning in Evo-devo, Evo-eco and Evolutionary...
50mFrom Classical Statistics to Modern Machine Learning