ChatPPG.
Paper

Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training

Richard Goldman, Varun Komperla, Thomas Ploetz, Harish Haresamudram
arXiv preprint2026IMU
Read paper arXiv Cite
Published articleFirst posted: Nov 8, 2025Publication date: Oct 10, 2026Added to atlas: Oct 12, 2026Source
Citations0Crossref referencesChecked Oct 16, 2026
GitHub stars—Not available
HF downloads—Not available

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

This paper tests zero-cost proxies, quick scores computed from one forward and backward pass, for picking good activity recognition architectures without training them. Across six datasets, the top-ranked architectures came within 7% of the best of 2,000 fully trained ones, and training just the top 10 came within 2%.

  • Evaluation: 6 HAR datasets, 2,000 sampled architectures
  • Result: top-10 picks within 2% of full-scale training
At a glance
Venue
arXiv preprint
Published
2026
DOI
10.1145/3830727.3834826 (opens in a new tab)
arXiv
2511.06157 (opens in a new tab)
PMID
Not reported
Licence
Not reported
Signals
IMU
Form factor
Not reported
Subjects
Not reported
Code
Not reported
Added
Oct 12, 2026

Abstract

Discovering high performing model architectures for wearables-based Human Activity Recognition (HAR) applications is challenging. The astonishing diversity and variability due to differing sensor locations, recording apparatus, activities, etc., can cause established architectures to perform worse on datasets/tasks they were not designed for. A promising complement to Neural Architecture Search (NAS) involves the development of Zero Cost Proxies (ZCPs), which correlate well with trained performance, but can be computed through a single forward/backward pass on a randomly sampled batch of data. In this paper, we investigate the effectiveness of eight ZCPs on six benchmark HAR datasets, and demonstrate that the top-predicted architectures obtain performance within 7% of that attained by full-scale training of 2,000 randomly sampled architectures. Furthermore, training the top-10 predicted architectures results in performance within 2% of full-scale training, leading to substantial computational savings. Our experiments introduce ZCPs to sensor-based HAR and demonstrate their suitability as an addition to NAS pipelines in practical scenarios.

Verbatim fromarXiv· arXiv metadata (abstracts are released under CC0)

Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training: ChatPPG research summary card

How it works

How Method works: Setup → Method → ResultSetup: IMU (Accelerometer, 3-axis). Method: Method, 8 zero-cost proxies for neural architecture search. Result: Evaluation (6 HAR datasets, 2,000 sampled architectures); Result (top-10 picks within 2% of full-scale training).SetupIMU: Accelerometer, 3-axisxyzIMUAccelerometer, 3-axisMethodMethod8 zero-cost proxies for neural…ResultEvaluation: 6 HAR datasets, 2,000 sampled architecturesEvaluation6 HAR datasets, 2,000 sampled…Result: top-10 picks within 2% of full-scale trainingResulttop-10 picks within 2% of full-scale…
Method
8 zero-cost proxies for neural architecture search

Cite

@article{modelsgottalentzerocostp2026,
  title = {{Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training}},
  author = {Richard Goldman and Varun Komperla and Thomas Ploetz and Harish Haresamudram},
  year = {2026},
  journal = {arXiv preprint},
  doi = {10.1145/3830727.3834826},
  eprint = {2511.06157}
}

History

  1. Added to the ChatPPG research hub

More on ChatPPG