Lately, I've been getting into energy-based models (EBMs) through some of Yann LeCun's recent talks, where he advocates the use of non-normalized models because it allows for more flexibility in the choice of the loss function and convenient inference over high-dimensional spaces.

However, after reading some papers on the recent approaches to training EBMs (e.g Kingma's How to Train Your Energy-Based Models), most approaches still use likelihood to optimize the EBMs parameters.

I'm lost in the necessity of using a normalized likelihood for training, while the whole idea of EBMs is that they are not normalized. Why are methods that shape the energy-function directly not popular?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.