Toward an AI Physicist for Unsupervised Learning

Abstract

We investigate opportunities and challenges for improving unsupervised machine learning using four common strategies with a long history in physics: divide-and-conquer, Occam's razor, unification and lifelong learning. Instead of using one model to learn everything, we propose a novel paradigm centered around the learning and manipulation of *theories*, which parsimoniously predict both aspects of the future (from past observations) and the domain in which these predictions are accurate. Specifically, we propose a novel generalized-mean-loss to encourage each theory to specialize in its comparatively advantageous domain, and a differentiable description length objective to downweight bad data and "snap" learned theories into simple symbolic formulas. Theories are stored in a "theory hub", which continuously unifies learned theories and can propose theories when encountering new environments. We test our implementation, the toy "AI Physicist" learning agent, on a suite of increasingly complex physics environments. From unsupervised observation of trajectories through worlds involving random combinations of gravity, electromagnetism, harmonic motion and elastic bounces, our agent typically learns faster and produces mean-squared prediction errors about a billion times smaller than a standard feedforward neural net of comparable complexity, typically recovering integer and rational theory parameters exactly. Our agent successfully identifies domains with different laws of motion also for a nonlinear chaotic double pendulum in a piecewise constant force field.

Replaced to match accepted PRE version. Added references, improved discussion. 22 pages, 7 figs

References in corpus (7)

• Overcoming catastrophic forgetting in neural networks
• Machine learning phases of matter
• Searching for Exotic Particles in High-Energy Physics with Deep Learning
• Learning phase transitions by confusion
• Discovering Phase Transitions with Unsupervised Learning
• Reconstructing quantum states with generative models
• A machine learns to predict the stability of circumbinary planets

Cited by in corpus (41)

• Social physics
• Discovering physical concepts with neural networks
• Machine Learning for Quantum Matter
• Deep symbolic regression for physics guided by units constraints: toward the automated discovery of physical laws
• Review: Deep Learning in Electron Microscopy
• AI Feynman 2.0: Pareto-optimal symbolic regression exploiting graph modularity
• Physics enhanced neural networks predict order and chaos
• Applying machine learning to improve simulations of a chaotic dynamical system using empirical error correction
• The Ramanujan Machine: Automatically Generated Conjectures on Fundamental Constants
• Symbolic Pregression: Discovering Physical Laws from Distorted Video
• The world as a neural network
• Falsification and consciousness
• Learning the Ising Model with Generative Neural Networks
• Learning Hamiltonian dynamics by reservoir computer
• Probabilistic Grammars for Equation Discovery
• Symbolic regression outperforms other models for small data sets
• Machine-Learning Non-Conservative Dynamics for New-Physics Detection
• Operationally meaningful representations of physical systems in neural networks
• Spectrally Adapted Physics-Informed Neural Networks for Solving Unbounded Domain Problems
• Modeling assembly bias with machine learning and symbolic regression
• Percent-level constraints on baryonic feedback with spectral distortion measurements
• Computers and turbulence
• Learning about learning by many-body systems
• Learning Generalized Quasi-Geostrophic Models Using Deep Neural Numerical Models
• Learning phase transitions from regression uncertainty: A new regression-based machine learning approach for automated detection of phases of matter
• Variational Autoencoder Analysis of Ising Model Statistical Distributions and Phase Transitions
• Renormalized Mutual Information for Artificial Scientific Discovery
• Neuronal diversity can improve machine learning for physics and beyond
• Conservation Laws and Spin System Modeling through Principal Component Analysis
• Using machine learning to find exact analytic solutions to analytically posed physics problems
• Computational Rational Engineering and Development: Synergies and Opportunities
• Discovering conservation laws from trajectories via machine learning
• Exploring fundamental laws of classical mechanics via predicting the orbits of planets based on neural networks
• Deep learning reveals hidden interactions in complex systems
• Celestial Machine Learning: Discovering the Planarity, Heliocentricity, and Orbital Equation of Mars with AI Feynman
• Boosting on the shoulders of giants in quantum device calibration
• Exploration and Exploitation in Symbolic Regression using Quality-Diversity and Evolutionary Strategies Algorithms
• Similarity-Based Equational Inference in Physics
• An Information Theory Approach to Physical Domain Discovery
• Toward Building Science Discovery Machines
• Embodied observations from an intrinsic perspective can entail quantum dynamics