ViViT: Curvature access through the generalized Gauss-Newton's low-rank structure

Curvature in form of the Hessian or its generalized Gauss-Newton (GGN)\napproximation is valuable for algorithms that rely on a local model for the\nloss to train, compress, or explain deep networks. Existing methods based on\nimplicit multiplication via automatic differentiation or Kronecker-factored\nblock diagonal approximations do not consider noise in the mini-batch. We\npresent ViViT, a curvature model that leverages the GGN's low-rank structure\nwithout further approximations. It allows for efficient computation of\neigenvalues, eigenvectors, as well as per-sample first- and second-order\ndirectional derivatives. The representation is computed in parallel with\ngradients in one backward pass and offers a fine-grained cost-accuracy\ntrade-off, which allows it to scale. We demonstrate this by conducting\nperformance benchmarks and substantiate ViViT's usefulness by studying the\nimpact of noise on the GGN's structural properties during neural network\ntraining.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC