Paper nostalgia series revisits ICLR 2020 classic: Your Classifier Is Secretly an Energy Based Model
CatAstro_Piyush · x · 2026-10-27
Researcher hmdolatabadi launched a "paper nostalgia" series to revisit the classics that made him fall in love with the field during his PhD. The first entry is Will Grathwohl et al.'s ICLR 2020 paper "Your Classifier is Secretly an Energy Based Model".
The paper reinterprets a standard discriminative classifier p(y|x) as an energy-based model over the joint distribution p(x,y) (JEM): class probabilities, unnormalized p(x) and p(x|y) all fall out of one framework, and training can leverage unlabeled data. Energy-based training improves calibration, robustness and out-of-distribution detection, enables samples rivaling GANs of the time, and adds little overhead over standard classification training.
More from Research
- Harmonic Launches Aristotle for Automated Formal Verification of C and Verilog Code — satnam6502 · 2026-10-28
- Verifiers have bugs too: flaws found in QED prover and RISC-V Lean model — xennygrimmato_ · 2026-10-28
- figwig data format generalizes bigwig to n-dimensional arrays for genomics — jmschreiber91 · 2026-10-28
- Infini-AI-Lab open-sources Hermit: continual learning via reusable adapters, no model changes — udmrzn · 2026-10-28
- MiRL: RL from recurring mistakes matches full terminal-agent training with 35% of rollouts — mohitban47 · 2026-10-28
- EvilQwen Reproduced: Neuron Steering as llama.cpp Control Vectors Survives Quantization — dahara111 · 2026-10-28