PapersSelected

Yann LeCun 澄清 EBM 常见误解:EBM 比概率模型更通用

Why we picked it

LeCun 亲自下场纠偏,很多人把 EBM 当概率模型的子集,其实是反过来的。搞自监督学习或研究 JEPA 架构的人值得花两分钟看完,理解为什么训练可以完全绕开配分函数。

Yann LeCun 在 X 上发文澄清能量模型(EBM)的一个常见误解:EBM 常被当作概率模型的一种特殊形式,但实际关系正好相反,EBM 是更一般的框架,概率模型只是其特例。

Body

他给出两个判定条件:一是能量能通过 Gibbs 分布转为概率,这要求配分函数收敛且可计算;二是损失函数有概率解释,如负似然。

LeCun 强调存在大量非概率的 EBM:它们不需要配分函数或归一化,可以用多种没有概率解释的损失函数训练。这一澄清与他长期主张的 JEPA 类非生成式架构方向一致,对理解自监督学习中的能量视角有帮助。

Original · Yann LeCun

IMPORTANT NOTE: EBM are often presented as a particular form of probabilistic models. This is a very common misunderstanding. EBMs are *MORE GENERAL* than probabilistic models. Probabilistic models are a special case of EBMs in which: 1. Energies can be turned into a probabilities through the Gibbs distribution. That requires a convergent and computable partition function! 2. The loss function has a probabilistic interpretation, e.g. negative likelihood. But there is whole world of EBMs that are NOT probabilistic: they require no partition function or normalization, and they can be trained with lots of loss functions that have no probabilistic interpretation.📊 11 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include EBM, Yann LeCun, energy-based models, reasoning models.

Source attribution: curated from Yann LeCun (@ylecun).