Reproduce “Variational inference for Monte Carlo objectives” 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
Andriy Mnih, Danilo J. Rezende
VIMCO provides an unbiased gradient estimator for training proposals in multi-sample objectives with discrete latent variables. It assigns each sample a learning signal based on the others, enabling sample-specific credit assignment. On binarized MNIST, a three-layer sigmoid belief network trained with 50 samples achieved test negative log-likelihood of 91.9 nats, compared with 96.2 for NVIL in this setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.