Equilibrium Training of Energy-Based Models with Parallel Trajectory Tempering
The paper claims Parallel Trajectory Tempering makes equilibrium maximum-likelihood training practical for EBMs.
The authors say poor MCMC mixing has kept energy-based models unreliable on difficult scientific datasets. Their PTT method uses the continuity of the training path to keep sampling near equilibrium during learning, with cost comparable to Persistent Contrastive Divergence. In RBM experiments, they report stronger performance than existing EBM training methods, plus equilibrium samples, thermalization-time estimates, and accurate log-likelihoods without much extra cost. On discrete tabular data, they say it beats state-of-the-art deep generative models, especially under limited data and overfitting pressure. ArXiv · AI/CL/LG's note
The authors say poor MCMC mixing has kept energy-based models unreliable on difficult scientific datasets. Their PTT method uses the continuity of the training path to keep sampling near equilibrium during learning, with cost comparable to Persistent Contrastive Divergence. In RBM experiments, they report stronger performance than existing EBM training methods, plus equilibrium samples, thermalization-time estimates, and accurate log-likelihoods without much extra cost. On discrete tabular data, they say it beats state-of-the-art deep generative models, especially under limited data and overfitting pressure. ArXiv · AI/CL/LG's note
score 4