We have all seen the demo: a computer vision model achieves 99% accuracy on a test dataset. Then, we deploy it to an edge device — a drone, a security camera, or an industrial robot — and performance crashes.
The problem is domain shift. The lighting is different, the camera angle is skewed, or the background noise has changed. In traditional MLOps, the solution is to collect thousands of new images from the edge device, label them manually, and retrain the model from scratch.
Read More from DZone.com Feed