hard negative mining

Hard negative mining is a method used when training AI where you deliberately pick wrong answers that look very similar to the correct one and show them to the model. Think of studying for an exam: instead of practicing random problems, you build a notebook of mistakes you keep making — especially ones where you confuse similar concepts — and drill on those over and over.

This approach matters because an AI trained only on obviously wrong answers stays weak in real-world use. For example, an AI learning to spot cat photos will quickly figure out the difference if you only show it car photos as wrong answers. But if you mix in photos of lynxes or dogs, which look much more like cats, the AI learns to make far more precise distinctions. The process of deliberately seeking out these hard, confusing wrong answers and feeding them into training is called hard negative mining.

This technique is especially important for building AI that needs to catch subtle differences between similar things — such as detectors that separate real search results from fake ones, face recognition systems, and document retrieval systems.