Hard Negatives

Computer programs that learn from examples need two kinds of examples. They need correct cases and they need incorrect cases. The incorrect cases are called negatives in technical jargon, that is, counterexamples. Most counterexamples are easy to see through: a system learning to recognize cats has no problem with a picture of a truck. Hard negatives are the difficult kind. These are counterexamples that look confusingly similar to the correct case — such as a fox instead of a cat. It is precisely on such cases that it is decided whether a system has really understood something.

Why easy counterexamples teach nothing

A learning system always looks for the most convenient path. If all counterexamples are obviously wrong, it finds a cheap shortcut. For cats and trucks, it may be enough to pay attention to fur and color. The system is almost always right during training and looks very good. In the real world, it then breaks down as soon as a dog appears in the picture.

You can imagine this like a class test in which all the wrong answers are obviously nonsense. An A grade on such a test says nothing about actual ability. Only when the wrong answers sound plausible does the test measure genuine knowledge. The same principle applies to training data.

That is why it often helps more to deliberately collect difficult counterexamples than to simply multiply the amount of data tenfold. Ten million boring examples sometimes barely improve a model. A few thousand well-chosen hard negatives shift the error rate noticeably. Here, data quality beats data quantity.

How to collect difficult counterexamples

The usual approach is called hard negative mining, that is, prospecting for hard cases. First, an initial version of the model is trained on normal data. Then this model is used to sift through large amounts of data. Wherever it confidently gets it wrong, there lies a hard negative. These cases are collected and used to train the next round.

With search systems, the process is similar. You ask a question and look at the hits the system ranks highly. The wrong ones among them are valuable because they are topically close. A text about interest rate cuts is a hard counterexample to a question about interest rate hikes. It shares almost all the words with the correct answer while meaning the opposite.

There is a trap here that often strikes in practice. Some supposedly wrong hits are in truth correct, just not marked as correct. If the system is trained to reject them, it learns something wrong. Experts then speak of false negatives. That is why mined hard negatives are frequently checked again, either by humans or by a second model.

Hard negatives in search, face recognition, and chatbots

Hard negatives are most visible in search engines and in the product search of online shops. Someone searching for wireless headphones should not get wired headphones. The two products differ in their description by only a few words. Such pairs are deliberately fed into training so that the system takes the difference seriously.

The principle is also central to face and speech recognition. Two siblings look more similar to each other than two random people. A recognition system trained only with random pairs will reliably confuse them. The same applies to fraud detection at banks: the dangerous cases are the ones that look like normal transfers.

In news about AI, the term usually appears in connection with retrieval, that is, with systems that fetch documents suitable for a chatbot. Companies then report that they have improved their search quality through better hard negative mining. What is almost always meant is the same thing: it’s not that the model got bigger, but that the training examples got meaner.