Bounding the Test Log-Likelihood of Generative Models
This addresses the challenge of assessing model quality for researchers and practitioners working with complex generative models, offering incremental improvements over prior methods.
The paper tackles the problem of evaluating generative models with intractable likelihoods by proposing a more efficient non-parametric density estimator that provides a lower bound on test log-likelihood and an unbiased estimator as sample size increases, with a biased variant for reliable model comparison using finite samples.
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.