Researchers have found a way to skip most of the training when hunting for the best neural network to power a wearable activity tracker.
Building a good AI model for wearable human activity recognition (HAR) is tricky because sensor placement, hardware, and the activities being tracked vary wildly between devices and datasets. A model architecture that works well on a smartwatch accelerometer might flop on a chest-strap sensor. Neural Architecture Search (NAS) can find the right fit, but training thousands of candidate architectures from scratch is slow and expensive. The researchers tested eight "zero-cost proxies" - scoring methods that estimate how well an architecture will perform after just one forward and backward pass on a batch of data, no full training required - across six benchmark HAR datasets.
The payoff: architectures picked by these proxies performed within 7% of what you'd get by fully training 2,000 randomly sampled architectures. Train just the top 10 proxy-picked candidates, and the gap shrinks to 2%, at a fraction of the compute cost. That matters for a field where every new dataset and sensor setup effectively demands a fresh architecture search.
Zero-cost proxies aren't new - they've sped up NAS for image and language models for a few years. What's notable here is that they hold up on the messier, noisier signal wearable sensors produce, where architecture rankings often don't transfer cleanly from other domains. It's a shortcut, not a replacement for training - a 7% gap is still a gap - but for a task where compute is usually the bottleneck, that's a shortcut worth taking.