کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Improved Generative Adversarial Network with Differentiable KS Distance

نویسندگان :

Siavash Sadeghi Ivrigh¹ Mohammadreza Hassannejad Bibalan² Asghar Keshtkar³

1- دانشگاه بین‌المللی امام خمینی (ره) 2- دانشگاه بین‌المللی امام خمینی (ره) 3- دانشگاه بین‌المللی امام خمینی (ره)

کلمات کلیدی :

GANs،KS distance،Sigmoid activation function،Deep learning،Machine learning

چکیده :

Generative adversarial networks (GANs) have emerged as a powerful framework for generating realistic data. However, evaluating the quality of generated samples remains a challenging task. This paper introduces a novel metric for evaluating GANs based on the Kolmogorov-Smirnov (KS) distance. The KS distance is a robust statistical measure to assess the similarity between two probability distributions. To efficiently compute the KS distance, we propose a novel approach that leverages the sigmoid activation function in the final layer of the discriminator. By carefully tuning the weights of this layer, we can approximate the KS distance between the real and generated data distributions. This approach offers a more efficient and effective way to evaluate GAN performance, leading to improved training and better-quality generated samples.