What are Energy-Based Models (EBMs) in unsupervised learning?

EBMs learn a scalar energy function where low energy corresponds to valid data configurations.

Energy-Based Models assign an energy score to input configurations, where real data has low energy and invalid data has high energy. Learning involves shaping this energy landscape without explicit labels, often using contrastive divergence or noise-contrastive estimation. EBMs are powerful because they do not require normalized probability distributions, but training is unstable due to partition function estimation challenges.