[Paper Review] Score-Guided Generative Adversarial Networks

Minhyeok Lee, Junhee Seok|arXiv (Cornell University)|Apr 9, 2020

Anomaly Detection Techniques and Applications32 references4 citations

TL;DR

This paper proposes Score-Guided GAN (ScoreGAN), a novel GAN architecture that uses a pre-trained evaluator network to guide generator training via the Inception score, avoiding overfitting to the Inception network. By employing the Inception score as an auxiliary loss rather than a primary target, ScoreGAN achieves state-of-the-art performance on CIFAR-10 (IS: 10.36 ± 0.15) and outperforms existing methods on CIFAR-100 (FID: 13.98).

ABSTRACT

We propose a Generative Adversarial Network (GAN) that introduces an evaluator module using pre-trained networks. The proposed model, called score-guided GAN (ScoreGAN), is trained with an evaluation metric for GANs, i.e., the Inception score, as a rough guide for the training of the generator. By using another pre-trained network instead of the Inception network, ScoreGAN circumvents the overfitting of the Inception network in order that generated samples do not correspond to adversarial examples of the Inception network. Also, to prevent the overfitting, the evaluation metrics are employed only as an auxiliary role, while the conventional target of GANs is mainly used. Evaluated with the CIFAR-10 dataset, ScoreGAN demonstrated an Inception score of 10.36$\pm$0.15, which corresponds to state-of-the-art performance. Furthermore, to generalize the effectiveness of ScoreGAN, the model was further evaluated with another dataset, i.e., the CIFAR-100; as a result, ScoreGAN outperformed the other existing methods, where the Fréchet Inception Distance (FID) was 13.98.

Motivation & Objective

• To address the issue of GANs overfitting to the Inception network when using its features for evaluation metrics.
• To improve sample quality and training stability in GANs by using an external pre-trained network as a score-guided evaluator.
• To demonstrate that auxiliary use of evaluation metrics like Inception score can enhance GAN performance without compromising generalization.
• To validate the generalization of the proposed method on diverse datasets beyond CIFAR-10, including CIFAR-100.

Proposed method

• Introduces a score-guided GAN (ScoreGAN) that uses a pre-trained classifier (e.g., Inception or other networks) as an evaluator module.
• Employs the Inception score as a regularization signal during generator training, rather than as the primary loss objective.
• Uses the pre-trained network to compute feature-based scores that guide the generator toward producing high-quality, diverse samples.
• Maintains the standard GAN training objective (e.g., adversarial loss) as the primary training signal, with the evaluation metric acting as a secondary auxiliary loss.
• Applies the method to both CIFAR-10 and CIFAR-100 datasets using the same architecture and training protocol.
• Ensures the evaluator network is fixed during training to prevent distribution shift and overfitting to adversarial examples.

Experimental results

Research questions

• RQ1 Can using a pre-trained network to guide GAN training via the Inception score improve sample quality and reduce overfitting?
• RQ2 Does treating the Inception score as an auxiliary loss rather than a primary objective lead to better generalization in GANs?
• RQ3 Can the score-guided training paradigm be generalized across different datasets, such as CIFAR-100?
• RQ4 How does ScoreGAN compare to state-of-the-art GANs in terms of Inception score and Fréchet Inception Distance?
• RQ5 What is the impact of using a non-Inception pre-trained network as the evaluator on avoiding adversarial overfitting?

Key findings

• ScoreGAN achieves an Inception score of 10.36 ± 0.15 on CIFAR-10, representing state-of-the-art performance at the time of publication.
• On CIFAR-100, ScoreGAN achieves a Fréchet Inception Distance (FID) of 13.98, outperforming existing methods.
• The use of the Inception score as an auxiliary loss reduces overfitting to the Inception network compared to standard GANs.
• The model generalizes well to other datasets, demonstrating consistent performance improvements over baseline GANs.
• The pre-trained evaluator network remains effective and stable throughout training, avoiding distributional shifts caused by adversarial examples.
• The method maintains high-quality sample diversity and fidelity, as evidenced by both Inception score and FID metrics.