Glossary¶
Definitions of the terms used throughout the ALF documentation.
- Active learning¶
A sequential strategy that, given a fixed labelling budget, repeatedly chooses the most informative or highest-value candidates to label next, rather than labelling everything or a random sample. ALF implements active learning as an ask/tell loop.
- Acquisition function¶
The rule that scores candidates by their expected value to the search and decides which to pick next, trading off exploitation (high predicted value) against exploration (high uncertainty). Built-in examples: Greedy, UCB, Expected Improvement, Thompson Sampling, CoreSet. Distinct from a
Search function, which produces the pool the acquisition function scores.- Best-found¶
The value of the best candidate discovered so far, tracked as a function of round. The headline “are we winning?” curve for a design experiment (see
Task).- Calibration¶
How well a model’s stated uncertainty matches its actual error. A well-calibrated
Surrogate“knows what it doesn’t know”, which is what lets anAcquisition functionexplore intelligently. Often summarised byECE.- Candidate¶
A single point in the search space (e.g. a protein sequence or a SMILES string), optionally carrying features and a label. Represented by
Candidate/LabelledCandidatesinalf_core.- Coverage¶
The fraction of the truly high-value region that an experiment has discovered. Distinguishes finding the peak from finding all the good designs; related to
Recall.- Dataset¶
The candidate pool together with its labels and train/validation splits, plus a query interface. Base class
BaseDataset; examples include GFP, FLIP, ProteinGym, and GuacaMol.- ECE¶
Expected Calibration Error: a scalar summary of
Calibration, the average gap between predicted confidence and observed accuracy across confidence bins.- Modality¶
The kind of a candidate — used to match datasets with compatible
Models and to pick a metric, not how the data is stored. It is the data’s domain where one exists (SEQUENCE,MOLECULE), withTABULARthe domain-agnostic case for raw numeric feature vectors. Storage type (used for serialisation) is inferred separately fromtype(data).- Model¶
A learnable predictor with a common interface:
featurise,train,predict,sample. Base classBaseModel; examples include CNN, Gaussian Process, ESM-2, and Chemprop. A model becomes aSurrogateor anOracledepending on the role it plays in the loop.- Offline¶
A loop in which the
Oraclereads labels from a held-out, pre-scored pool. Fully reproducible with no external calls; the mode used for benchmarking and method development.- Online¶
A loop in which the
Oracleobtains labels from a live scorer (a trained model or simulator) on demand; the mode used to drive a real experiment. The loop is otherwise identical toOffline.- Optimizer¶
The object that bundles an
Acquisition functionand aSearch functionand exposes theask(propose a batch) andtell(retrain on new labels) interface. ClassOptimizer.- Oracle¶
The source of ground-truth labels for a proposed batch. Wraps a scorer that is either a trained
Model(online) or aDataset(offline). ClassOracle.- Precision¶
Of the candidates a method selected as high-value, the fraction that truly are. Reported alongside
Recallfor classification-style objectives.- Recall¶
Of all the truly high-value candidates, the fraction the experiment has selected. Measures whether the method is finding the good region, not just one point; see also
Coverage.- Regret¶
The gap between the best achievable value and the best value found so far. Lower is better; regret-versus-round is the primary way ALF compares the speed of methods.
- Search function¶
The component that generates the pool of
Candidateobjects anAcquisition functionthen scores, e.g. enumerating a dataset (DatasetSearch) or sampling from a generativeModel(GeneratorSearch). Base classBaseSearch.- Spearman correlation¶
A rank-correlation metric used to judge how well a
Surrogate’s predicted ordering of candidates matches the true ordering: what matters for selection, regardless of absolute scale.- State¶
The mutable carrier threaded through the loop: the current
Dataset,Surrogate, round index, acquisition history, and per-round metrics. ClassState.- Surrogate¶
A cheap-to-evaluate probabilistic approximation of the true objective, wrapping a
Modeland retrained each round on the data acquired so far. Its predictions (and uncertainty) drive theAcquisition function. ClassSurrogate.- Task¶
The orchestrator that runs the loop end to end via
setupthenrun. The family of task (DesignTask,SupervisedTask, orZeroShotTask) determines what is being optimised. Base classBaseTask.