Hybrid Weakly Self-Supervised and Multi-Task Representation Learning for Label-Efficient Human Activity Recognition Using Wearable Sensors

Abstract

Wearable Human Activity Recognition (HAR) has taken a center stage in intelligent health monitoring, smart environments, and context-aware computer systems as a key component. In spite of the current developments in the field of deep learning, a majority of the state-of-the-art HAR systems require large amounts of labeled data, which restricts their scalability and application to a variety of real-life conditions. Sensor streams are costly, time-consuming, and unreliable to be labeled as this method is expensive and time consuming, there is noisiness in labeling as well as user variability. In this paper, a hybrid learning model that combines weakly self-supervised learning and multi-task representation learning to allow recognition of human activities using weak labelling of multimodal wearable sensors is proposed. The strategy builds on the auxiliary learning goals based on unlabeled sequences as well as correlated representation extraction based on the correlated predictive tasks. The framework performance is based on a combination of self-directed feature discovery and task-directed inductive bias to promote generalization, decrease reliance on manual labeling, as well as increase resilience to incomplete/noisy labeling. The methodology focuses on the temporal signal modeling, multimodal sensor fusion and adaptive optimization of joint loss goals. It is expected to yield better quality representation, competitive classification performance as compared to fully supervised baselines and scaling learning behavior to be applicable to large scale deployment. The paper helps to develop faster data-aware machine intelligence in wearable computing systems and gives hints on hybrid learning frameworks that can be used in contexts with larger time-series analytics.