Reproduce “GFlowNet-EM for learning compositional latent variable models” 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
Edward J. Hu, Nikolay Malkin, Moksh Jain, Katie Everett, Alexandros Graikos, Yoshua Bengio
GFlowNet-EM enables maximum-likelihood estimation for latent variable models with discrete, compositional latents by leveraging GFlowNets to approximate the intractable E-step of the EM algorithm. The framework achieved improved performance on hierarchical mixture models, grammar induction, and discrete VAEs by learning expressive, non-factorized posterior distributions.
Key insights
- Generative Flow Networks (GFlowNets) can be effectively integrated into the EM framework to perform amortized inference for intractable posterior distributions over discrete combinatorial latents.
- This integration allows for maximum-likelihood estimation of latent variable models capable of learning expressive, non-factorized posteriors, addressing a limitation of traditional variational EM.
- Techniques like a "sleep phase" and GFlowNet-guided MCMC significantly improve the quality and mode coverage of the approximate posterior, mitigating common issues like posterior collapse.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.