Abstract
Knowledge distillation (KD) has been actively studied for image classification tasks in deep learning, aiming to improve the performance of a student model based on the knowledge from a teacher model. However, there have been very few efforts for applying KD in image regression with a scalar response, and there is no KD method applicable to both tasks. Moreover, existing KD methods often require a practitioner to carefully choose or adjust the teacher and student architectures, making these methods less scalable in practice. Furthermore, although KD is usually conducted in scenarios with limited labeled data, very few techniques are developed to alleviate such data insufficiency. To solve the above problems in an all-in-one manner, we propose in this paper a unified KD framework based on conditional generative adversarial networks (cGANs), termed cGAN-KD. Fundamentally different from existing KD methods, cGAN-KD distills and transfers knowledge from a teacher model to a student model via cGAN-generated samples. This unique mechanism makes cGAN-KD suitable for both classification and regression tasks, compatible with other KD methods, and insensitive to the teacher and student architectures. Also, benefiting from the recent advances in cGAN methodology and our specially designed subsampling and filtering procedures, cGAN-KD also performs well when labeled data are scarce. An error bound of a student model trained in the cGAN-KD framework is derived in this work, which theoretically explains why cGAN-KD takes effect and guides the implementation of cGAN-KD in practice. Extensive experiments on CIFAR-10 and Tiny-ImageNet show that we can incorporate state-of-the-art KD methods into the cGAN-KD framework to reach a new state of the art. Also, experiments on RC-49 and UTKFace demonstrate the effectiveness of cGAN-KD in image regression tasks, where existing KD methods are inapplicable.
Abstract (translated)
URL
https://arxiv.org/abs/2104.03164
