PDD: Pruning Neural Networks During Knowledge Distillation

Although deep neural networks have developed at a high level, the large computational requirement limits the deployment in end devices. To this end, a variety of model compression and acceleration techniques have been developed. Among these, knowledge distillation has emerged as a popular approach that involves training a small student model to mimic the performance of a larger teacher model. However, the student architectures used in existing knowledge distillation are not optimal and always have redundancy, which raises questions about the validity of this assumption in practice. This study aims to investigate this assumption and empirically demonstrate that student models could contain redundancy, which can be removed through pruning without significant performance degradation. Therefore, we propose a novel pruning method to eliminate redundancy in student models. Instead of using traditional post-training pruning methods, we perform pruning during knowledge distillation (PDD) to prevent any loss of important information from the teacher models to the student models. This is achieved by designing a differentiable mask for each convolutional layer, which can dynamically adjust the channels to be pruned based on the loss. Experimental results show that with ResNet20 as the student model and ResNet56 as the teacher model, a 39.53%-FLOPs reduction was achieved by removing 32.77% of parameters, while the top-1 accuracy on CIFAR10 increased by 0.17%. With VGG11 as the student model and VGG16 as the teacher model, a 74.96%-FLOPs reduction was achieved by removing 76.43% of parameters, with only a loss of 1.34% in the top-1 accuracy on CIFAR10. Our code is available at https://github.com/YihangZhou0424/PDD-Pruning-during-distillation.