NOLS: A Near-sensor On-chip Learning System with Direct Feedback Alignment for Personalized Wearable Heart Health Monitoring

Detecting cardiac arrhythmia is crucial in preventing heart attacks, and wearable electrocardiograph (ECG) systems have been developed to address this issue. However, the typical 'first off-chip learning, then on-chip processing' strategy poses significant challenges in practicality for personalized edge systems. In this paper, we first propose a near-sensor on-chip learning and inference system with direct feedback alignment for user-specific cardiac arrhythmia detection. This system features an event-driven near-sensor feature extraction module and a hybrid on-chip learning and inference processor. Through system-level co-design, our proposed on-chip learning solution achieves almost lossless classification performance with an accuracy of 98.56%, which is among the best. Compared to backpropagation on GPU, our approach only incurs less than 0.5% accuracy loss. Additionally, a configurable processor architecture is proposed and verified, supporting parallel learning and pipelined inference to reduce both energy consumption and system latency.