Efficient human activity classification via sparsity-driven transfer learning

The deployment of deep neural networks in real-world radar-based human activity classification is largely hindered by both the high computational cost and the large amount of training samples. In this study, the authors propose a method to simultaneously reduce the computational burden and the number of labelled training samples. Different from previous transfer learning methods that simply prune fully-connected layers and modify the weights of the convolutional layers, they enforce filter-level sparsity in the transfer learning from ImageNet to the micro-Doppler measurements. Through the sparsity-driven transfer learning, unimportant convolutional filters can be identified and then be pruned. Therefore, a light but effective transfer learned net can be obtained. The experiments demonstrate the sparsity-driven transfer learned VGG-19 Net not only outperforms convolutional neural networks trained from scratch by nearly 10% accuracy but also gives an 11 $ \times $× reduction in the number of parameters and a 10 $ \times $× reduction in computing operations compared with the original VGG-19 Net.