Layer-Wise Training Convolutional Neural Networks With Smaller Filters for Human Activity Recognition Using Wearable Sensors

Recently, convolutional neural networks (CNNs) have set latest\nstate-of-the-art on various human activity recognition (HAR) datasets. However,\ndeep CNNs often require more computing resources, which limits their\napplications in embedded HAR. Although many successful methods have been\nproposed to reduce memory and FLOPs of CNNs, they often involve special network\narchitectures designed for visual tasks, which are not suitable for deep HAR\ntasks with time series sensor signals, due to remarkable discrepancy.\nTherefore, it is necessary to develop lightweight deep models to perform HAR.\nAs filter is the basic unit in constructing CNNs, it deserves further research\nwhether re-designing smaller filters is applicable for deep HAR. In the paper,\ninspired by the idea, we proposed a lightweight CNN using Lego filters for HAR.\nA set of lower-dimensional filters is used as Lego bricks to be stacked for\nconventional filters, which does not rely on any special network structure. The\nlocal loss function is used to train model. To our knowledge, this is the first\npaper that proposes lightweight CNN for HAR in ubiquitous and wearable\ncomputing arena. The experiment results on five public HAR datasets, UCI-HAR\ndataset, OPPORTUNITY dataset, UNIMIB-SHAR dataset, PAMAP2 dataset, and WISDM\ndataset collected from either smartphones or multiple sensor nodes, indicate\nthat our novel Lego CNN with local loss can greatly reduce memory and\ncomputation cost over CNN, while achieving higher accuracy. That is to say, the\nproposed model is smaller, faster and more accurate. Finally, we evaluate the\nactual performance on an Android smartphone.\n