Towards real-world continual learning for human activity recognition

Sensor-based human activity recognition (HAR) is essential for applications in health monitoring, smart homes, and manufacturing. However, the long-term deployment of HAR systems presents a significant challenge: how to automatically evolve a HAR model to adapt to changes in the environment or activity patterns. This thesis addresses this issue by introducing an online continual learning (OCL) scenario for HAR, where sensor data streams contain unlabeled samples from both known and new activities. We propose OCL-HAR, a technique that performs real-time predictions on streaming sensor data while simultaneously discovering new activities. Empirical evaluations on four HAR datasets demonstrate that the OCL scenario is challenging for state-of-the-art continual learning (CL) techniques. In contrast, OCL-HAR consistently outperforms these techniques by up to 17% and 23% in micro and macro F1 scores, respectively. Additionally, on-device CL is crucial for deploying applications on battery-powered devices, balancing privacy needs and adaptability. However, existing CL techniques can be cost-prohibitive for such devices, requiring quantised operations with reduced precision for practical deployment. Current fully quantized training (FQT) solutions fail to converge in low-precision CL. To address this, we propose hadamard domain quantized training (HDQT), which utilises the Hadamard transform to enable 4-bit FQT with integer operands. HDQT facilitates low-precision, on-device training where other FQT solutions fail. HDQT achieves better gradient alignment with unquantised baselines in early feature detection layers, leading to consistently improved learning performance, as reflected in loss landscapes trajectories. This research highlights the potential of OCL-HAR and HDQT for efficient, real-time, and low-power CL in practical HAR applications.