Active learning

What it is

Most machine learning starts with data someone else has already gathered. Active learning turns that around. The model is given a small amount of data to begin with, builds a rough picture of the problem, and then picks the next thing to measure. The measurement comes back, the model updates, and it picks again. Each pass round this loop is a cycle, and the rule it uses to decide what to try next is called an acquisition rule.

A good rule balances two urges. One is to measure where the model already expects a good result, to confirm it. The other is to measure where the model is most unsure, because that is where it has most to learn. Models that report their own uncertainty, rather than only a single answer, suit this well. When the apparatus can run several samples at once, the model picks a batch instead of one.

Why scientists use it

Experiments cost money, time and material, and the space of things that could be tried is usually far larger than anyone can get through. A chemist cannot cook every mixture; a telescope cannot watch every flickering dot in the sky with a spectrograph. Choosing well therefore matters more than measuring fast.

Active learning stands in for the judgement of an experienced researcher deciding what to try tomorrow in the light of today. Its advantage is that it never tires, keeps the whole record in view, and can be wired directly into equipment so the choosing and the doing happen without a pause. That pairing of model and instrument is what people mean by a self-driving or closed-loop laboratory.

In the record

structural-biologyRobot lab and learning agent redesign an enzyme to survive higher temperaturesBefore running experiments, several ways of choosing the next protein sequences to measure, including random picking, were compared on an existing enzyme dataset in simulated trials.materials-chemistryRobotic lab with machine learning proposes and runs 353 inorganic synthesis experimentsSynthesis targets were screened from stability calculations, then precursors and a predicted temperature were supplied for each, with one temperature per target so recipes could be batched.astronomyTelescope targets chosen nightly by a classifier learning from its own spectraStarting from thirty labelled light-curves, a model gave each telescope alert a probability of being an early supernova, to decide which deserved spectroscopic follow-up.materials-chemistryA learning algorithm ran a synchrotron beamline to find a phase-change materialAt each cycle of the run, measured diffraction patterns were sorted into phase regions by a method encoding physical constraints, informed by spectra taken beforehand.

Limits

A loop can only explore what it was told to consider, so a promising region left out of the list stays invisible. It trusts its own estimate of uncertainty, and if that estimate is poor it may keep sampling the same comfortable corner and call the result the best there is. It also inherits every fault of its instrument: a drifting sensor or a quietly failing step teaches the model something false, and because the model then chooses the next experiment, the error compounds rather than averaging out.

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