Reproduce “Learning the Stein Discrepancy for Training and Evaluating Energy-Based Models without Sampling” 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
Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Richard Zemel
Researchers from the University of Toronto and Vector Institute introduce Learned Stein Discrepancy (LSD), a framework for training and evaluating energy-based models without relying on costly and biased Markov Chain Monte Carlo sampling. LSD uses a neural network to estimate Stein Discrepancy, demonstrating superior performance in high-dimensional hypothesis testing and model evaluation, and enabling stable training of EBMs on image datasets.
Key insights
- Stein Discrepancy, when estimated by a neural critic with efficient trace approximation, provides a robust, sampler-free, and scalable method for EBM analysis.
- Relying only on first-order derivatives for both the model and critic contributes to numerical stability, especially in high-dimensional learning tasks.
- A unified framework allows for consistent training and rigorous evaluation of unnormalized density models, mitigating the biases of MCMC-dependent approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.