Feature Affinity Assisted Knowledge Distillation and Quantization of Deep Neural Networks on Label-Free Data
This work addresses the practical challenge of model compression for scenarios where labeled data is scarce or expensive, offering an incremental improvement over existing knowledge distillation techniques.
The paper tackles the problem of compressing deep neural networks via quantization-aware training by introducing a feature affinity assisted knowledge distillation method that works on label-free data, achieving stronger supervision for the student network than from labeled ground-truth data.
In this paper, we propose a feature affinity (FA) assisted knowledge distillation (KD) method to improve quantization-aware training of deep neural networks (DNN). The FA loss on intermediate feature maps of DNNs plays the role of teaching middle steps of a solution to a student instead of only giving final answers in the conventional KD where the loss acts on the network logits at the output level. Combining logit loss and FA loss, we found that the quantized student network receives stronger supervision than from the labeled ground-truth data. The resulting FAQD is capable of compressing model on label-free data, which brings immediate practical benefits as pre-trained teacher models are readily available and unlabeled data are abundant. In contrast, data labeling is often laborious and expensive. Finally, we propose a fast feature affinity (FFA) loss that accurately approximates FA loss with a lower order of computational complexity, which helps speed up training for high resolution image input.