Fast model screening for wearable activity trackers cuts training time by 99%

Finding the right neural network architecture for wearable-based activity recognition is computationally expensive, but a new approach could dramatically accelerate the process. Researchers have applied zero-cost proxies (ZCPs)—lightweight evaluation metrics computed in a single forward and backward pass—to identify promising models without training them from scratch.

The Challenge in Wearables#

Human activity recognition systems powered by wearable sensors face a fundamental problem: the vast diversity of sensor placements, recording equipment, and activity datasets means that architectures optimized for one task often fail on others. This variability has made it difficult to establish universally reliable network designs, leaving practitioners to either reuse existing models (which may underperform) or conduct expensive architecture searches.

Zero-Cost Proxies as a Solution#

Rather than training thousands of candidate architectures to find the best one, the researchers tested eight different ZCPs—metrics that estimate model quality without full training—against six benchmark HAR datasets. The method screens architectures using only random initialization and a small batch of data, then ranks candidates by their proxy scores.

Results and Savings#

The findings show that architectures ranked highest by ZCPs achieve performance within 7% of full training across 2,000 randomly sampled models. More impressively, training just the top 10 predicted architectures yields results within 2% of exhaustive search—a significant computational win. This approach substantially reduces the search burden while maintaining competitive performance.

Implications for Practice#

This work introduces ZCPs to the sensor-based HAR domain and demonstrates their viability as a standard component in neural architecture search pipelines, especially for real-world deployments where computational budgets are tight. The method promises to make custom model discovery faster and more accessible for wearables applications across different use cases.