Symbolic Pregression: Discovering Physical Laws from Raw Distorted Video

Silviu-Marian Udrescu

MIT Dept. of Physics & Center for Brains, Minds & Machines

A central goal of physics and science more broadly is to discover mathematical patterns in data. We present a method for unsupervised learning of exact equations of motion for objects in raw and optionally distorted unlabeled videos. We first train an autoencoder that maps each video frame into a low-dimensional latent space where the laws of motion are as simple as possible. Differential equations describing the motion are then discovered using Pareto-optimal symbolic regression. We find that our pre-regression (“pregression") step is able to rediscover Cartesian coordinates of unlabeled moving objects even when the video is distorted by a generalized lens. Our approach represents a first step towards building AI systems able to distill physics information from complex systems, without any human intervention.