Answers are grounded in this paper and cite their sources.
The rapid proliferation of Wireless Fidelity (Wi-Fi) sensing technology faces significant challenges related to trustworthiness and robustness, particularly when deployed in unpredictable environments with noisy labels. Although deep learning approaches have achieved high accuracy, they often lack physical interpretability and may fail catastrophically under distributional shifts. To address these limitations, we introduce PhiFi, a physics-informed deep learning framework that integrates electromagnetic principles-specifically, the Helmholtz wave equation-through a supervised Physics Head and Residual Gating mechanism for device-free human activity recognition. Our framework improves calibration consistency, achieving an Expected Calibration Error (ECE) of 2.84% on synthetic data (SRV protocol, compared to 3.39% for purely data-driven baselines). On real-world UT-HAR benchmark data, calibration reaches 4.1% (LOSO cross-domain protocol), matching the baseline's cross-domain generalization (≈84% macro-F1) while demonstrating improved label efficiency in extreme scarcity regimes (1% labels: +25.7 pp improvement under STEA protocol). Comprehensive evaluations across synthetic stress conditions and real benchmark datasets indicate that physics-informed constraints provide a practical approach for improving model trustworthiness and calibration, offering potential benefits for safety-critical Internet of Things (IoT) applications in data-limited scenarios.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: