Over-Cleaned Training Data Can Break Production Vision Models

Machine learning teams often spend weeks removing blurry, underexposed, and occluded images to build pristine training datasets. Such datasets can produce strong validation scores but then degrade once models are deployed, failing under glare, rain, and camera vibration. The article argues that over-sanitizing strips out the structured real-world noise models need to learn robust representations. It recommends cleaning only destructive errors such as corrupted frames and label misalignments while preserving physical variations like motion blur, sensor glare, and occlusion. Teams are advised to write an explicit Data Preservation List before any cleaning effort begins.