Speaker
Ms
Anindita Maiti
(Northeastern University)
Description
Untrained asymptotically wide Neural Networks are Gaussian Processes, with a direct correspondence to Euclidean free field theory; deviations of the NN away from GP can be effectively described using Wilsonian EFT. Further, output dimension d shows up as the number of independent species in the free / interacting field corresponding to GP / non-GP NN outputs. Experimentally, we verify Ward identity for n-point correlation functions of Gauss-net architecture in both GP and NGP limit in appropriate regime of validity of EFT description. Resulting SO(d) symmetry has interesting consequences on perturbative corrections to GP correlation functions, in terms of couplings corresponding to interaction terms of EFT action.
