Towards optimal sparse CNNs: sparsity-friendly knowledge distillation through feature decoupling

Despite the efficacy of network sparsity in reducing the complexity of convolutional neural networks (CNNs), the performance of sparse networks often deteriorates significantly compared to their dense counterparts. Knowledge distillation is regarded as a potent strategy for utilizing large models to augment the performance of smaller counterparts; however, its advantages for sparse networks remain substantially constrained. We identify in this article that the underlying issue stems from sparse student models exhibiting disparate behaviors in processing foreground and background features, thereby hindering the uniform transfer of knowledge from dense models that address both feature types concurrently. Building on this insight, we introduce a novel sparsity-friendly knowledge distillation (SF-KD) method, which independently supervises the two feature types using feature decoupling to facilitate effective knowledge distillation for sparse networks. Specifically, we decouple the foreground and background features through unique pooling techniques and implement separate mean squared error (MSE) feature distillation. Furthermore, we dynamically adjust the weights of the two loss components to optimize performance. Experimental results on Canadian Institute For Advanced Research (CIFAR) datasets (including CIFAR-10 and CIFAR-100) and Mini-ImageNet benchmarks substantiate significant performance enhancements, underscoring the effectiveness of our proposed methodology.