arXiv:2511.06157v3 Announce Type: replace-cross
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.
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.
This article has been indexed from cs.AI updates on arXiv.org
Read the original article: