Reproduce “Natural Reweighted Wake-Sleep” 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
Csongor Várady, Riccardo Volpi, Luigi Malagò, Nihat Ay
NRWS trains Helmholtz Machines to model handwritten-digit distributions with generator and recognition networks. It replaces ordinary gradient updates with natural-gradient updates, using conditional independence to make the Fisher information matrix block diagonal by neuron. On binarized MNIST, its test log-likelihood estimate was −84.91, versus −87.36 for RWS and −86.88 for diagonal-FIM DNRWS. Convergence was faster, but epochs took longer in this setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.