Block-Wise Training Residual Networks on Multi-Channel Time Series for Human Activity Recognition

Recently, human activity recognition (HAR) has become an active research area in wearable computing scenario. On the other hand, residual nets have continued to push the state-of-the-art of computer vision and natural language processing. However, residual nets have rarely been considered in the HAR field. As residual nets grow deeper, memory footprint limit its wide use for a variety of HAR tasks. In this paper, we present a novel block-wise training residual nets that use local loss functions for HAR applications. Instead of global backprop, the local cross-entropy loss together with a supervised local similarity matching loss is utilized to train each residual block independently, in which gradient need not to be propagated down the network. As a result, the gradient and activations do not have to be kept in memory any more, which alleviates the memory requirements and is more beneficial for wearable HAR computing. We demonstrate the effectiveness of block-wise training residual nets on OPPORTUNITY, WISDM, UNIMIB SHAR and PAMAP2 datasets, which establishes obvious better classification accuracy compared to equally-sized residual nets, even though memory requirement is much smaller.