Neural Nets: Beyond Equilibrium Physics

Neural networks offer a unique optimization approach that tracks entire trajectories and non-equilibrium evolution, differentiating their application in fundamental physics from traditional variational principles.

• Neural network optimization tracks full trajectories, unlike traditional physics focusing on minima/maxima.

• This 'out of equilibrium' evolution in ML systems can explain classical behavior emergence.

• Physicists use machine learning as a computational tool, not yet widely as a model for physical systems.

Neural networks use a variational principle, but unlike traditional physics, they are interested in the entire evolutionary trajectory rather than just minimum or maximum states. This 'out of equilibrium' evolution, prevalent in optimizing machine learning systems, allows for the emergence of classical-like behavior. While physicists use machine learning as a computational tool for complex problems, its distinct optimization approach suggests its potential as a model for fundamental physics.

Neural Net Optimization

• 00:00:00 Neural network optimization differs from traditional physics by focusing on the entire trajectory from an initial state to a complicated final state, not just the extrema of an action. This approach tracks the 'out of equilibrium' evolution inherent in machine learning, which is not typically present in classical physics. This allows for the emergence of classical-like behaviors through learning and optimization processes.

ML as Physics Tool

• 00:01:30 Physicists frequently adopt machine learning as a powerful computational tool to solve complex problems, such as finding the ground state of a quantum many-body system. However, this application is primarily as a tool for computation rather than as a fundamental model of a physical system itself. The optimization methods are leveraged for their utility in tackling difficult calculations.