The recent advent of powerful generative models has triggered the renewed development of quantitative measures to assess the proximity of two probability distributions. As the scalar Frechet Inception Distance remains popular, several methods have explored computing entire curves, which reveal the trade-off between the fidelity and variability of the first distribution with respect to the second one. Several of such variants have been proposed independently and while intuitively similar, their relationship has not yet been made explicit. In an effort to make the emerging picture of generative evaluation more clear, we propose a unification of four curves known respectively as: the Precision-Recall (PR) curve, the Lorenz curve, the Receiver Operating Characteristic (ROC) curve and a special case of Rényi divergence frontiers. In addition, we discuss possible links between PR / Lorenz curves with the derivation of domain adaptation bounds.

In this work, we propose a new parameter efficient sharing method for the training of GAN generators. While there has been recent progress in transfer learning for generative models with limited data, they are either limited to domains close to the original one, or adapt a large part of the parameters. This is somewhat redundant, as the goal of transfer learning should be to reuse old features. In this way, we propose width wise parameter sharing, which can learn a new domain with ten times fewer trainable parameters without a significant drop in quality. Previous approaches are less flexible than our method and also fail to preserve image quality for challenging transfers. Finally, as our goal is ultimately parameter reuse, we show that our method performs well in the multi-domain setting, wherein several domains are learned simultaneously with higher visual quality than the state of the art StarGAN-V2.

With the widespread application of deep networks in industry, membership inference attacks, i.e. the ability to discern training data from a model, become more and more problematic for data privacy. Recent work suggests that generative networks may be robust against membership attacks. In this work, we build on this observation, offering a general-purpose solution to the membership privacy problem. As the primary contribution, we demonstrate how to construct surrogate datasets, using images from GAN generators, labelled with a classifier trained on the private dataset. Next, we show this surrogate data can further be used for a variety of downstream tasks (here classification and regression), while being resistant to membership attacks. We study a variety of different GANs proposed in the literature, concluding that higher quality GANs result in better surrogate data with respect to the task at hand.

State of the art deep generative networks have achieved such realism that they can be suspected of memorizing training images. It is why it is not uncommon to include visualizations of training set nearest neighbors, to suggest generated images are not simply memorized. We argue this is not sufficient and motivates studying overfitting of deep generators with more scrutiny. We address this question by i) showing how simple losses are highly effective at reconstructing images for deep generators ii) analyzing the statistics of reconstruction errors for training versus validation images. Using this methodology, we show that pure GAN models appear to generalize well, in contrast with those using hybrid adversarial losses, which are amongst the most widely applied generative methods. We also show that standard GAN evaluation metrics fail to capture memorization for some deep generators. Finally, we note the ramifications of memorization on data privacy. Considering the already widespread application of generative networks, we provide a step in the right direction towards the important yet incomplete picture of generative overfitting.

In this article we revisit the definition of Precision-Recall (PR) curves for generative models proposed by (Sajjadi et al., 2018). Rather than providing a scalar for generative quality, PR curves distinguish mode-collapse (poor recall) and bad quality (poor precision). We first generalize their formulation to arbitrary measures hence removing any restriction to finite support. We also expose a bridge between PR curves and type I and type II error (a.k.a. false detection and rejection) rates of likelihood ratio classifiers on the task of discriminating between samples of the two distributions. Building upon this new perspective, we propose a novel algorithm to approximate precision-recall curves, that shares some interesting methodological properties with the hypothesis testing technique from (Lopez-Paz & Oquab, 2017). We demonstrate the interest of the proposed formulation over the original approach on controlled multi-modal datasets.