Search papers, labs, and topics across Lattice.
Department of Mathematics, University of Pisa, Italy
2
0
3
Finite-width neural networks converge to their Gaussian-process limits with error bounds that shrink as the network width increases, revealing a surprising robustness across architectures.
Bayesian neural networks actually learn features in the large width limit, defying the conventional wisdom of fixed-kernel NNGP theory.
