Why architecture matters here

Active learning fails on cold start + biased strategy. Architecture matters because query + human + retrain compose.

Advertisement

The architecture: every piece explained

The top strip is loop. Unlabeled pool. Model. Query strategy. Select batch.

The middle row is process. Human labeler. Labeled set. Retrain. Budget.

The lower rows are ops. Diversity + representativeness. Metrics. Ops — QA + cold start + stop rule.

Active learning — query strategy + human loop + budgetselect samples to label to maximize model gainUnlabeled poollargeModelcurrentQuery strategyuncertainty / marginSelect batchtop-kHuman labelerannotatorsLabeled setgrowingRetraineach cycleBudgetfixed labels/dayDiversity + representativenessnot just uncertainMetricslabel efficiency curveOps — annotator QA + cold start + stop rulelabelgrowretrainbudgetdiversewatchwatchoperateoperate
Active learning loop: model queries samples to label.
Advertisement

End-to-end flow

End-to-end: current model predicts on 100k unlabeled. Uncertainty ranks them; top-500 sent to labelers with diversity constraint. Labeled set grows. Retrain. Label efficiency curve tracked; stop when plateau.