DeActive

Abstract

Deep learning architectures have been applied increasingly in multi-modal problems which has empowered a large number of application domains needing much less human supervision in the process. As unlabeled data are abundant in most of the application domains, deep architectures are getting increasingly popular to extract meaningful information out of these large volume of data. One of the major caveat of these architectures is that the training phase demands both computational time and system resources much higher than shallow learning algorithms and it is posing a difficult challenge for the researchers to implement the architectures in low-power resource constrained devices. In this paper, we propose a deep and active learning enabled activity recognition model, DeActive, which is optimized according to our problem domain and reduce the resource requirements. We incorporate active learning in the process to minimize the human supervision along with the effort needed for compiling ground truth. The DeActive model has been validated using real data traces from a retirement community center (IRB #HP-00064387) and 4 public datasets. Our experimental results show that our model can contribute better accuracy while ensuring less amount of resource usages in reduced time compared to other traditional deep learning approaches in activity recognition.