Data Cleansing for GANs
Naoyuki Terashita, Hiroki Ohashi, Satoshi Hara
TL;DR
The paper tackles the problem of improving GAN performance by identifying and removing harmful training instances. It develops a Jacobian-based influence framework that captures indirect effects between the discriminator and generator, introducing two estimators, ITD and AID, to approximate the impact of removing a sample without full retraining. It then extends this approach to quantify influence on GAN evaluation metrics via ITD-EIGEM and AID-EIGEM, enabling data cleansing that improves metrics such as Inception Score, FID, and ALL across diverse GAN architectures (LQGAN, DCGAN, StyleGAN, StyleGAN2 with LoRA). The findings demonstrate that removing identified harmful instances yields consistent gains and that one-epoch retraining can be nearly as effective as full retraining for cleansing, offering practical computational advantages. Overall, the work provides a principled, scalable route to enhance generative performance by judicious data pruning grounded in cross-model influence analysis.
Abstract
As the application of generative adversarial networks (GANs) expands, it becomes increasingly critical to develop a unified approach that improves performance across various generative tasks. One effective strategy that applies to any machine learning task is identifying harmful instances, whose removal improves the performance. While previous studies have successfully estimated these harmful training instances in supervised settings, their approaches are not easily applicable to GANs. The challenge lies in two requirements of the previous approaches that do not apply to GANs. First, previous approaches require that the absence of a training instance directly affects the parameters. However, in the training for GANs, the instances do not directly affect the generator's parameters since they are only fed into the discriminator. Second, previous approaches assume that the change in loss directly quantifies the harmfulness of the instance to a model's performance, while common types of GAN losses do not always reflect the generative performance. To overcome the first challenge, we propose influence estimation methods that use the Jacobian of the generator's gradient with respect to the discriminator's parameters (and vice versa). Such a Jacobian represents the indirect effect between two models: how removing an instance from the discriminator's training changes the generator's parameters. Second, we propose an instance evaluation scheme that measures the harmfulness of each training instance based on how a GAN evaluation metric (e.g., Inception score) is expected to change by the instance's removal. Furthermore, we demonstrate that removing the identified harmful instances significantly improves the generative performance on various GAN evaluation metrics.