Reproduce “Combined Representation and Generation with Diffusive State Predictive Information Bottleneck” 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
Richard John, Yunrui Qiu, Lukas Herron, Pratyush Tiwary
A new framework, Diffusive State Predictive Information Bottleneck (D-SPIB), combines physics-inspired representation learning with score-based generative diffusion models to create a unified system for analyzing and generating molecular configurations. It demonstrates enhanced generative performance and the ability to accurately infer and extrapolate thermodynamic behavior across varying temperatures, even from limited training data.
Key insights
- Jointly optimizing a physically-motivated representation (SPIB) and a generative model (diffusion) in a single framework is more effective than separate approaches.
- Replacing a simple analytical prior in SPIB with a flexible, learned prior from a diffusion model substantially improves the accuracy of reconstructing complex latent space distributions.
- Embedding temperature information directly into the diffusion process enables the model to effectively infer and extrapolate thermodynamic properties from limited multi-temperature data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.