Active learning

Active learning is a training strategy where the model helps decide which data points to label next, rather than labeling everything at random. The idea is simple: if the model is already confident about certain images, labeling more of those adds little value. Instead, the model flags images where it's most uncertain, and human annotators focus their effort there. This gets you better performance with fewer labeled examples.

Common selection strategies include uncertainty sampling (pick images where the model's top prediction confidence is lowest), query-by-committee (train multiple models and pick images where they disagree most), and diversity sampling (pick images that are most different from what's already labeled to cover more of the data distribution). In practice, teams combine these approaches for best results.

Active learning is especially valuable when labeling is expensive, which covers most computer vision tasks. Medical imaging requires expert radiologists, industrial inspection requires domain engineers, and even general object detection takes 10-60 seconds per bounding box. An active learning loop can cut labeling costs by 40-60% compared to random annotation while reaching the same model accuracy.