Training with an answer key

Look at what the correction depends on. The machine guesses dog. Something compares that guess with the known answer, cat, and the difference guides the adjustment.

The answer must already be attached to the example before this kind of training can use it.

People might label the examples directly. The answers might also come from existing records, such as whether a past payment was fraudulent.

This arrangement is called . The supervision is the answer key.

It has a plain cost. Useful answers must exist for enough examples, and those answers must be reliable.

People do some of that work under poor conditions. Investigations have documented low-paid workers labelling graphic and harmful text so later systems could recognize it. This is one part of data work, not the story of every training dataset. Someone had to read it returns to this at the end of the course.