Few-shot transfer learning for wearable IMU-based human activity recognition

Deep learning has proven to be highly effective for human activity recognition (HAR) when large amount of labelled data is available for the target task. However, training a deep learning model to generalize well on a new task with just-few observations of labelled data is an active area of research. In this paper, a novel few-shot transfer learning (FSTL) approach is proposed for classification of human activities using just few instances (shots) of the data obtained from a wearable system assembled to collect inertial sensor data for different human activities, performed by two users. First, a deep learning model is trained on a large publicly available HAR dataset. The model parameters of such a model are then fine-tuned using the Reptile algorithm to determine the optimal initial parameter set using which, the model will classify activities with just few-shots of data from the target task. The proposed FSTL approach yields an average classification accuracy of 74.86 ± 0.71% and 79.20 ± 1.05% for 3-way, 5-shot classification of new activities performed by a single user and same set of activities performed by a new user, respectively. When the pre-trained weights are used as the initial weights in the Reptile algorithm, the generalization ability of the model improves by about 10% for 3-way, 5-shot classification as compared to using few-shot learning without parameter transfer.