Machine Learning, Kolmogorov Complexity, and Squishy Bunnies

Abstraction: Using neural networks to approximate physics simulation via Kolmogorov complexity intuition and PCA

Key points:

• Kolmogorov complexity: length of the shortest program that produces a function's outputs; low-complexity functions are good candidates for neural network approximation
• Neural networks lie on a computation-memory diagonal: both scale proportionally with number of weights, and accuracy decreases with function complexity
• PCA applied to physics simulation data extracts major axes of deformation and approximates the function's Kolmogorov complexity (number of axes needed)
• Cloth/soft-body physics simulations are less complex than they appear; 256 PCA axes can reconstruct ~3000-vertex cloth motion adequately
• Resulting neural network is 300-5000x faster than direct simulation (35-350 μs/frame), enabling real-time use in game production (Ubisoft research)

Connections: Ubisoft · Kolmogorov Complexity · Neural Networks · Physics Simulation · Principal Component Analysis

Source: http://www.theorangeduck.com/page/machine-learning-kolmogorov-complexity-squishy-bunnies