Experiment

What it is

In research that uses artificial intelligence, a great deal happens inside a computer. A model predicts, a program searches, a statistical method sorts data into groups. An experiment is the step that none of that can replace. Something is physically done: a chemical is weighed out and heated, a protein is made in a test tube and then chopped at by an enzyme, a telescope is pointed at a patch of sky and its light split into a spectrum. The result is a new fact about the world rather than a new inference from facts already held.

Experiments can be run by hand at a bench, or by machines. A robot can dose powders, stir them and put them in a furnace. An instrument can be told by software which sample to measure next. When the deciding and the doing are joined into one cycle, with software choosing the next experiment and equipment carrying it out without a person in between, the arrangement is often called an autonomous or self-driving laboratory. Without a physical step there is nothing for such a laboratory to drive.

Why scientists use it

Scientists use experiments because prediction and measurement are different things. A model trained on past results can only reproduce the patterns in what it was given; it has no way to find out whether a substance that nobody has yet made will hold together, or whether a point of light in the sky is the kind of exploding star it appears to be. Only doing the thing settles that.

Experiments also supply the raw material that models need. The measurements a laboratory produces become the data on which the next model is trained, and the data against which its guesses are checked. In an automated setup the experiment plays a second role as well: it is the slow, costly, finite step. Computation is cheap and can be repeated endlessly, while a furnace run or a night of telescope time cannot. That scarcity is usually what the surrounding software is there to manage, by choosing which few physical measurements are worth making.

In the record

Limits

An experiment answers only the question actually put to the apparatus, under the conditions used. It can fail for mundane reasons, such as a badly mixed powder or cloud over a telescope, and a failure of that kind looks much like a genuine negative result. Automation makes the doing fast but does not make it correct: a robot repeats a flawed procedure faithfully, and an instrument under software control will dutifully measure whatever it is pointed at. What can be measured also shapes what gets studied, quietly narrowing the questions asked.

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

Every article