Joint energy-based models (JEM) are a class of neural networks that handle both classification and generation within a single architecture, and they also support out-of-distribution (OOD) detection. This dual capability makes them an attractive target for research, but training them is not straightforward.

A new paper on arXiv examines the training procedure for canonical JEMs. The standard approach relies on stochastic gradient Langevin dynamics (SGLD) to sample from the model's distribution. The authors propose and evaluate a fixed-noise predictor-corrector adaptation as an alternative sampling method.

The comparison is carried out on the CIFAR-10 dataset, a common benchmark for image classification and generation. The paper's abstract does not state the outcome of the comparison, so the relative merits of the two methods remain to be seen from the full text.