Satoshi Sugiyama

Observable- and Positional-Encoding-Dependent Symmetry Readout from Neural Network Weights

Naoya Chiba, Satoshi Sugiyama, Yuki Uranishi

Post-hoc analysis of trained neural network weights often seeks to recover geometric structure directly from the parameters. We show that, for positional-encoding-equipped neural fields, the symmetry visible from weights is not the true symmetry group itself, but an observable sy…Read more