Reproduce “Directly Training Joint Energy-Based Models for Conditional Synthesis and Calibrated Prediction of Multi-Attribute Data” 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
Jacob Kelly, Richard Zemel, Will Grathwohl
This work introduces Gibbs-JEM, an extension of Joint Energy-Based Models (JEMs) that directly trains the joint distribution of multi-attribute data and labels using a Langevin-Within-Gibbs sampling approach. The model achieves superior calibration performance for multi-attribute predictions, notably an ECE of 0.2503 on UTZappos50K, and successfully synthesizes high-quality data conditioned on specified attribute combinations, including those not present in training.
Key insights
- Directly training the joint distribution in Energy-Based Models enables more flexible and direct conditional synthesis compared to approaches with factorized likelihoods.
- The Langevin-Within-Gibbs sampling algorithm effectively addresses the challenges of sampling from mixed continuous and discrete distributions within EBMs.
- A single EBM can simultaneously achieve strong discriminative performance with calibrated predictions and powerful generative capabilities for multi-attribute data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.