EBM tutorial thread: ML learning methods differ only in negative sampling

daibond_alpha · x · 2026-11-20

Part 2/5 of the EBM tutorial thread: every EBM is three choices — energy (what it represents), inference (how to sample/optimize), learning (how to fit). MLE, CD, pseudolikelihood, score matching, NCE, and adversarial training differ only in how negatives are sampled (q=p₀, k MCMC steps, one Gibbs step, one Langevin step, weighted noise, learned sampler), with a visual comparison.

Original post →

More from Research

Research channel →