News & Events

Subscribe to email list

Please select the email list(s) to which you wish to subscribe.

User menu

You are here

Improved conditional generative adversarial networks for image generation: methods and their application in knowledge distillation

Tuesday, August 31, 2021 - 11:00 to 12:00
Xin Ding, UBC Statistics PhD Student

To Join Via Zoom: To join this seminar, please request Zoom connection details from headsec [at]

Abstract: Conditional generative adversarial networks (cGANs) are state-of-the-art models for synthesizing images dependent on some conditions. These conditions are usually categorical variables such as class labels. cGANs with class labels as conditions are also known as class-conditional GANs. Some modern class-conditional GANs such as BigGAN can even generate photo-realistic images. The success of cGANs has been shown in various applications. However, two weaknesses of cGANs still exist. First, image generation conditional on continuous, scalar variables (termed regression labels) has never been studied. Second, low-quality fake images still appear frequently in image synthesis with state-of-the-art cGANs, especially when training data are limited. This thesis aims to resolve the above two weaknesses of cGANs and explore the applications of cGANs in improving a lightweight model with the knowledge from a heavyweight model (i.e., knowledge distillation).

First, existing empirical losses and label input mechanisms of cGANs are not suitable for regression labels, making cGANs fail to synthesize images conditional on regression labels. To solve this problem, this thesis proposes the continuous conditional generative adversarial network (CcGAN), including novel empirical losses and label input mechanisms.

Moreover, even the state-of-the-art cGANs may produce low-quality images, so a subsampling method to drop these images is necessary. In this thesis, we propose a density ratio based subsampling framework for unconditional GANs. Then, we introduce its extension to the conditional image synthesis setting called cDRE-F-cSP+RS, which can effectively improve the image quality of both class-conditional GANs and CcGAN.

Finally, we propose a unified knowledge distillation framework called cGAN-KD suitable for both image classification and regression (with a scalar response), where the synthetic data generated from class-conditional GANs and CcGAN are used to transfer knowledge from a teacher net to a student net, and cDRE-F-cSP+RS is applied to filter out bad-quality images. Compared with existing methods, cGAN-KD has many advantages, and it achieves state-of-the-art performance in both image classification and regression tasks.