Reproduce “The continuous Bernoulli: fixing a pervasive error in variational autoencoders” 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
Gabriel Loaiza-Ganem, John P. Cunningham
The authors show that a grayscale-image VAE using a Bernoulli likelihood is not properly normalized for pixel intensities in [0,1]. They train it instead with the continuous Bernoulli, whose normalization changes the reconstruction objective. On original MNIST, the continuous-Bernoulli-trained model scores about 300 nats higher on the continuous-Bernoulli ELBO than the Bernoulli-trained model, under the reported experiments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.