Reproduce “Improving black-box optimization in VAE latent space using decoder uncertainty” 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
Pascal Notin, José Miguel Hernández-Lobato, Yarin Gal
Researchers developed a method to enhance black-box optimization in Variational Autoencoder (VAE) latent spaces by quantifying decoder epistemic uncertainty using an importance sampling-based estimator, consistently yielding more valid and higher-quality generated objects across diverse domains.
Key insights
- The IS-MI estimator consistently provides more robust and lower-variance estimates of epistemic uncertainty in structured data compared to naive Monte Carlo or token-independence approximations.
- Regions of the latent space where the VAE decoder exhibits low epistemic uncertainty are highly correlated with the generation of valid and high-quality discrete objects.
- Guiding latent space optimization by avoiding high-uncertainty regions enables a better trade-off between maximizing the black-box objective and ensuring the validity and quality of the generated outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.