On-Sensor Online Learning and Classification Under 8 KB Memory

Inertial sensors provide a low-power and high-fidelity pathway for state estimation and sensor fusion. Inertial measurement units now feature on-chip processors for ultra-low-power information fusion, signal processing, and neural network-based classification at the extreme edge. However, accounting for domain shifts, personalized inference requirements, and application diversity makes adopting existing learning-enabled on-device training, classification, and fusion frameworks for on-sensor processors difficult. This paper introduces a method for personalized and on-device learning for on-chip classification, inference, and information fusion applications. The proposed framework automatically segments and stores quantized gravity vector image templates and axes variance information of motion artifacts during training. During inference, templates created from the time-series windows are matched against uniform blurred templates using the universal image quality index. An adaptive rep counting module robust to varying motion primitives counts repetitions of matched motion primitives. The framework requires no human-engineered parameters and allows for the personalization and addition of new motion artifacts. Our framework recognizes human activities with 96.7% test accuracy and achieves an average rep count error of 0.44, while reducing the memory usage by 1000-2000$\times$ over existing tiny machine learning on-device learning techniques, allowing on-sensor learning and inference under 8 KB of memory.