Self-gated recurrent neural networks for human activity recognition on wearable devices

This paper develops a self-gated recurrent neural network (SGRNN), and applies it to human activity recognition (HAR), using timeseries signals collected from embedded sensors of wearable devices. Recurrent neural networks (RNNs) are very powerful for timeseries signal analysis. Especially, by integrating gates into recurrent units, gated RNNs such as LSTM and GRU are more complexity, and do not suffer from the vanishing gradient problem, so can learn very long-term dependencies. However, for use on wearable devices, RNNs must be simplified to reduce resource consumption, including memory usage and computational cost. The proposed model is approximately the same size and burdensome computation as that of a standard RNN, but exhibits explicit properties of the gating mechanism, so it is unaffected by the problem of vanishing gradients. Experimental results on the HAR problem not only demonstrate that the accuracy of our model is superior to that of the standard RNN, and is comparable with that of LSTM and GRU, but the model is low in resource consumption.