کنفرانس مهندسی برق ایران

صفحه اصلی / سی و دومین کنفرانس بین المللی مهندسی برق

Enhancing SCGAN’s Disentangled Representation Learning with Contrastive SSIM Similarity Constraints

نویسندگان :

Iman Yazdanpanah¹ Ali Eslamian²

1- دانشگاه صنعتی اصفهان 2- دانشگاه صنعتی اصفهان

کلمات کلیدی :

Generative Adversarial Nets،Unsupervised Learning،Disentangled Representation Learning،Contrastive Disentanglement،SSIM

چکیده :

SCGAN adds a similarity constraint between generated images and conditions as a regularization term on generative adversarial networks. Similarity constraint works as a tutor to instruct the generator network to comprehend the difference in representations based on conditions. We understand how SCGAN works on a deeper level. This understanding makes us realize that the similarity constraint functions like the contrastive loss function. Two major changes we applied to SCGAN to make a modified model are using SSIM to measure similarity between images and applying contrastive loss principles to the similarity constraint. The modified model performs better using FID and Factor metrics. On the MNIST and Fashion-MNIST datasets, the modified model achieves log-likelihood values of 234.8 and 332.6, respectively, surpassing SCGAN’s 232.5 and 324.2. The modified model exhibits improvements in FID, attaining values of 3.42 and 12.97 for MNIST and Fashion-MNIST, respectively, compared to SCGAN’s 4.11 and 14.63. In terms of disentanglement, the modified model shows clear advantages, achieving values of 0.89 and 0.91 for MNIST and Fashion-MNIST, respectively, compared to SCGAN’s 0.77 and 0.89. These improvements show that the modified model has effectively learned a more disentangled representation compared to SCGAN. The modified model also has better generalisability compared to other models.