Data-free Learning of Reduced-order Kinematics
DescriptionWe present a new approach for finding a small subspace that parameterizes the low-energy configurations of a physical system by "overfitting" a simple neural network. No dataset is needed, just the system's potential energy function. The method applies broadly to elastic deformables, cloth, rigid bodies, and even linkages.
TimeWednesday, 9 August 20232:55pm - 3:05pm PDT