Bo Dai and Yilun Du release EBM tutorial unifying ML through one energy-based lens
daibond_alpha · x · 2026-11-20
Bo Dai and Yilun Du published the tutorial "An Energy-Based Perspective on Machine Learning" (arXiv:2611.14383), arguing that EBMs are not just one model family but a single lens on all of ML.
- Every EBM is defined by three choices: the energy (what the model represents), inference (how to sample or optimize it), and learning (how to fit it).
- MLE, NCE, score matching, pseudolikelihood, and adversarial training share one contrastive structure, differing only in how negatives are sampled.
- Through this lens, discriminative models, GANs, flows, diffusion, autoregressive LMs, SimCLR/CLIP/word2vec/BERT, EBM policies, world models, IRL, planning, and inference-time computation (best-of-N, verifiers) are all EBMs differing mainly in partition-function and sampling handling.
- The framework explains existing methods and suggests new algorithms by swapping one component while keeping the others, plus principled composition of energies.
Related event: Bo Dai and Yilun Du Release Energy-Based Model Tutorial(2 posts)→
More from AGI Musings
- Lex Sokolin podcast: you can beat 98% of competitors and still get wiped out by the wave you missed — LexSokolin · 2026-11-20
- Investor bullish on Anthropic stock as one of few companies that reach RSI — _arohan_ · 2026-11-20
- Economist Daniel Susskind brings sensible AI talk to The Daily Show — joshgans · 2026-11-20
- Denny Zhou: since 2026, the best AI models make the best data — denny_zhou · 2026-11-20
- Do LLMs kill off-computer thinking? Two researchers disagree — Afinetheorem · 2026-11-20
- Emad Mostaque: 'human in the loop' breaks when AI outruns review — VraserX · 2026-11-20