Validation

What it is

A method that makes predictions has to be judged somehow. Validation does this by comparing its output with a reference: a measurement made by other means, a result worked out by a person, or an answer settled by earlier work. That reference is often called ground truth, meaning the best available account of what is actually the case. The comparison is only meaningful if the reference was not used to build or tune the method in the first place. Otherwise the method is being asked to repeat something it has already seen, and will look better than it is. So data is usually held back, kept aside and untouched, and brought out only at the end.

Validation needs a measure as well as a reference. Scientists count how often the method agrees, how much it is off by, how many real cases it catches and how many of the cases it flags turn out to be false. Some checks are not statistical at all. A prediction can be tested by going back to the bench and altering the thing predicted, to see whether the system behaves as the prediction implies. Others compare a fast automated result with a slow careful one done by hand on the same material. All of these ask the same question in different ways: does the output survive contact with something independent?

Why scientists use it

Modern research methods, especially those built by machine learning, are good at producing confident output. A model will return an answer for almost any input, and the answer carries no mark of whether it is sound. Validation is what stands in for the judgement of an expert who would have checked the work. It converts a plausible-looking result into a quantity: this often right, this often wrong, wrong in this particular way. Without it, a method's output is a suggestion rather than a finding.

It also tells scientists where a method may be used. A validated error rate sets how much weight a result can bear, whether it is enough on its own or needs confirmation by another technique. When an automated system runs without a person watching each step, validation is the only point at which human judgement enters, so it carries the whole weight of trust in what the system reports.

In the record

Limits

Validation only answers for the cases it was given. A method checked on well-behaved, well-studied examples may fail on the awkward ones, because references exist mostly where earlier work has already succeeded. The reference itself can be wrong, so agreement may mean two methods share a mistake. A single summary figure hides how errors are distributed, and a method can score well overall while failing badly on the cases that matter most. Validation also says nothing about why a method works, so it cannot warn when the conditions it was tested under quietly change.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Every article