TinyHAR-UQ: Battery-aware, uncertainty-controlled tinyML for wearable activity recognition on IoT edge devices

Abstract

Tiny, battery-powered wearables increasingly run human activity recognition (HAR) models, yet they rarely know when to trust predictions or when to stop computing to save energy. We address this gap with TinyHAR-UQ, a microcontroller-ready HAR pipeline that turns int8 networks into risk-aware, energy-aware predictors. Our contributions are threefold. First, we realize on-device conformal prediction sets for HAR by combining an int8 temporal backbone, early exits, Dirichlet (evidential) heads, and a streaming integer-only conformal layer that enforces user-chosen miscoverage levels. Second, we design a budget-aware controller that, at test time, decides whether to exit early, continue, or abstain based on prediction-set size and resource constraints. Third, we propose micro-temperature personalization, a zero-cost calibration scheme (two scalars per exit folded into quantized scales/biases) that recovers calibration lost to quantization. Deployed on Cortex-M4F and ESP32-S3 devices and evaluated on four HAR benchmarks, TinyHAR-UQ achieves nominal 90-95% coverage with near-singleton sets while cutting median latency and energy by up to 50% and 20-30% versus a matched single-exit baseline, enabling reliable, efficient HAR on off-the-shelf IoT edge platforms.