Knowledge Transferring in Deep Learning of Wearable Dynamics

Wearable dynamics learning is attracting intensive interests now in the era of smart health. One challenge is how the learning model can be effective trained with limited amounts of data. Targeting this, we in this study propose a deep transfer learning framework, to leverage the knowledge obtained from non-target users and boost the performance on the target-user. More specifically, we have designed and pretrained a convolutional neural network on the non-target database, and then fine-tuned the model on a small portion of the target database. The framework has been evaluated on a wearable biomechanical learning application for physical activity detection. Compared with direct target-data-based learning, the proposed deep transfer learning approach great boosts the detection accuracy. This study will advance the wearable dynamics learning applications through deep knowledge transferring.