EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

UASD-Edge: Uncertainty-Aware Self-Distillation for Early Human Activity Recognition on Edge Devices

Antti Saarinen, Laura Niemi, Matthias Peeters, Jonas Verhoeven

Journal of Clinical Practice and Health Sciences · 25.08.2026 · 6 (2) · 76-86

DOI: 10.71465/jcphs72 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Clinical Practice and Health Sciences, online erschienen am 25.08.2026. Verfasst von Antti Saarinen und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Early human activity recognition (HAR) on edge devices must trade classification quality against sensing delay while avoiding unreliable supervision from incomplete observations. This paper presents UASD-Edge, a compact early-HAR framework that augments full-window-to-prefix self-distillation with per-sample teacher reliability. The reliability is computed from normalized entropy of the detached full-window prediction and weights only the distillation term, leaving hard-label supervision unchanged. A single 7,995-parameter temporal CNN is shared by all prefix lengths, and no uncertainty branch or additional inference parameters are introduced. We evaluate on the official subject-disjoint UCI Human Activity Recognition Using Smartphones split using the three dynamic locomotion classes and six inertial sequence channels. Across seeds 13, 37, and 73, UASD-Edge obtains macro-F1 of 94.51±1.18%, 97.88±1.12%, 98.22±1.07%, and 98.56±0.77% at 25%, 50%, 75%, and 100% observation. Against a same-backbone PrefixKD baseline with the identical distillation coefficient, UASD is lower by 0.22 percentage points at 25% but improves the later three checkpoints by 0.14–0.19 points, so no statistical-significance claim is made. A validation-selected global MSP gate reaches 98.05±0.75% test accuracy and 98.03±0.76% macro-F1 after observing 32.59±2.96% of the 2.56-s window on average (0.834 s). Frozen TorchScript batch-1 FP32 inference takes 0.0319 ms median at the full window on one Intel Xeon Platinum 8573C CPU thread. The results support uncertainty-weighted distillation as a low-cost regularizer for later prefixes while exposing a reproducible accuracy-earliness operating point.

Zitieren

APA

Saarinen, A., Niemi, L., Peeters, M., & Verhoeven, J. (2026). UASD-Edge: Uncertainty-Aware Self-Distillation for Early Human Activity Recognition on Edge Devices. Journal of Clinical Practice and Health Sciences, 6(2), 76-86. https://doi.org/10.71465/jcphs72

Vancouver

Saarinen A, Niemi L, Peeters M, Verhoeven J. UASD-Edge: Uncertainty-Aware Self-Distillation for Early Human Activity Recognition on Edge Devices. Journal of Clinical Practice and Health Sciences. 2026;6(2):76-86. doi:10.71465/jcphs72

BibTeX
@article{Saarinen2026,
  author = {Saarinen, Antti and Niemi, Laura and Peeters, Matthias and Verhoeven, Jonas},
  title = {UASD-Edge: Uncertainty-Aware Self-Distillation for Early Human Activity Recognition on Edge Devices},
  journal = {Journal of Clinical Practice and Health Sciences},
  year = {2026},
  volume = {6},
  number = {2},
  pages = {76--86},
  doi = {10.71465/jcphs72}
}
Kostenlos starten →