English

Bounding the Test Log-Likelihood of Generative Models

Machine Learning 2014-05-13 v4

Abstract

Several interesting generative learning algorithms involve a complex probability distribution over many random variables, involving intractable normalization constants or latent variable normalization. Some of them may even not have an analytic expression for the unnormalized probability function and no tractable approximation. This makes it difficult to estimate the quality of these models, once they have been trained, or to monitor their quality (e.g. for early stopping) while training. A previously proposed method is based on constructing a non-parametric density estimator of the model's probability function from samples generated by the model. We revisit this idea, propose a more efficient estimator, and prove that it provides a lower bound on the true test log-likelihood, and an unbiased estimator as the number of generated samples goes to infinity, although one that incorporates the effect of poor mixing. We further propose a biased variant of the estimator that can be used reliably with a finite number of samples for the purpose of model comparison.

Keywords

Cite

@article{arxiv.1311.6184,
  title  = {Bounding the Test Log-Likelihood of Generative Models},
  author = {Yoshua Bengio and Li Yao and Kyunghyun Cho},
  journal= {arXiv preprint arXiv:1311.6184},
  year   = {2014}
}

Comments

10 pages, 1 figure, 2 tables. International Conference on Learning Representations (ICLR'2014, conference track)

R2 v1 2026-06-22T02:14:00.090Z