Session Name: | Machine Learning: Physics Simulation, Kolmogorov Complexity, and Squishy Bunnies |
Speaker(s): | Daniel Holden |
Company Name(s): | Ubisoft Montreal |
Track / Format: | Programming |
Did you know free users get access to 30% of content from the last 2 years?
Overview: | We know Machine Learning is a powerful tool to tackle problems we can't solve by conventional means, but what about things we already have solutions for, such as physics simulations? Is there any reason to use Machine Learning in these cases? It turns out there is because in some cases Machine Learning allows us to trade-off computation time in exchange for additional memory usage, which can often be used to provide massive performance gains at runtime. In this talk I will show how we used Neural Networks and vast amounts of training data to construct extremely fast approximations of interactive physics and cloth simulations which achieved around 300 to 5000 times speedup over standard simulations, opening up new opportunities for what is possible within typical physics simulation budgets. |