Reproduce “Slicing Mutual Information Generalization Bounds for Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kimia Nadjahi, Kristjan Greenewald, Rickard Brüel Gabrielsson, Justin Solomon
The authors make information-theoretic generalization bounds computable for neural networks by measuring how much training data is revealed by weights restricted to a frozen, randomly chosen lower-dimensional subspace. For models only approximately representable in that subspace, they add a distortion term. On MNIST and CIFAR-10, they report non-vacuous bounds while retaining test performance at suitable subspace dimensions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.