Reproduce “DVAE#: Discrete Variational Autoencoders with Relaxed Boltzmann Priors” 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

Arash Vahdat, Evgeny Andriyash, William G. Macready

DVAE# enables importance-weighted training for binary Boltzmann priors by smoothing binary latent variables and marginalizing them into a continuous prior. On statically binarized MNIST, its two-group nonlinear model with K=25 achieves 82.85±0.03, versus 84.19±0.21 for DVAE and 83.58±0.15 for DVAE++. These are test-set estimates, not exact likelihoods, and apply only to the tested setting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.