学术文献库

Measuring the generalization capacity of Deep Generative Models (DGMs) is difficult because of the curse of dimensionality. Evaluation metrics for DGMs such as Inception Score, Fréchet Inception Distance, Precision-Recall, and Neural Net Divergence try to estimate the distance between the generated distribution and the target distribution using a polynomial number of samples. These metrics are the target of researchers when designing new models. Despite the claims, it is still unclear how well can they measure the generalization capacity of a generative model. In this paper, we investigate the capacity of these metrics in measuring the generalization capacity. We introduce a framework for comparing the robustness of evaluation metrics. We show that better scores in these metrics do not imply better generalization. They can be fooled easily by a generator that memorizes a small subset of the training set. We propose a fix to the NND metric to make it more robust to noise in the generated data. Toward building a robust metric for generalization, we propose to apply the Minimum Description Length principle to the problem of evaluating DGMs. We develop an efficient method for estimating the complexity of Generative Latent Variable Models (GLVMs). Experimental results show that our metric can effectively detect training set memorization and distinguish GLVMs of different generalization capacities. Source code is available at https://github.com/htt210/GeneralizationMetricGAN.