Does active learning always ask for the most uncertain examples?

No. Strategies can also consider diversity, representativeness, expected improvement, or combinations of criteria.

Can a pretrained deep model be used?

Yes. The selection loop can use or adapt an existing representation; training from scratch is not required by the concept.

Is a fixed evaluation set still necessary?

Yes. It helps measure progress independently of the examples chosen by the acquisition strategy.

AI Glossary / Deep learning / Intermediate

Deep Active Learning

Definition

Deep active learning combines deep neural models with a process for selecting which examples should receive labels next. Instead of labeling an available collection indiscriminately, the system chooses candidates expected to help learning, obtains their labels, and updates the model. The aim is to use an annotation budget effectively. It does not guarantee that every selected example will improve performance or that the fewest theoretically possible labels will be found.

What it means

Imagine having thousands of unlabeled images but only enough expert time to label a small portion. The selection decision becomes part of the learning workflow. Which images would be most useful to label now, given what the model already knows and what the application needs?

“Deep” describes the model or representation involved. “Active” means that the learning system influences which information it requests. This differs from a model merely processing incoming data automatically. Human annotators often supply labels, but the defining feature is selective querying of a label source, not the presence of a particular annotation interface.

How it works

Begin with an initial labeled set and an unlabeled pool. Train or adapt a model, score candidate examples using an acquisition strategy, and select a batch for annotation. Add the resulting labels to the training collection and repeat within the available budget.

Strategies may emphasize uncertainty, disagreement between model predictions, or coverage of the representation space. Deep Bayesian active learning studies uncertainty-based selection; core-set approaches study selecting representative batches.

Keep an independent evaluation set. The acquisition pool is shaped by the selection process, so performance on queried examples alone does not provide an unbiased picture of the deployed task. Label quality and annotation cost also need attention.

Three practical examples

1. Visual inspection: an engineer selects a limited batch of uncertain and varied images for specialist annotation, then checks performance on a fixed representative evaluation set.

2. Document processing: a team chooses examples from underrepresented document layouts for review rather than repeatedly labeling near-duplicates from the most common layout.

3. Research comparison: students compare random sampling, uncertainty sampling, and a diversity-aware strategy at the same cumulative annotation budgets and with repeated initial seeds.

Think of it this way

Think of a student choosing questions to ask a tutor after identifying gaps in their understanding. A useful question can reveal more than another easy exercise, but asking only confusing or unanswerable questions may waste time. The analogy is limited: model uncertainty is a computed signal and may be poorly calibrated. The learner does not inherently know which examples are important or whether an annotator can label them reliably.

Worked example

Suppose a binary classifier gives four unlabeled examples probabilities of 0.50, 0.55, 0.90, and 0.99 for the positive class. Under a simple uncertainty rule based on closeness to 0.5, the first two examples would be selected if the budget allowed two labels.

Now suppose those two examples are nearly identical frames from the same video. Labeling both may provide less new information than labeling one of them plus a different example from another condition. This motivates considering batch diversity, although it does not identify the best strategy without evaluation.

For budget accounting, assume an initial set contains 100 labels and each of four rounds requests 25 more. The final labeled set contains 100 + 4 × 25 = 200 examples, assuming no duplicates or rejected queries.

To claim label savings, compare performance with a baseline at matched budgets and count the initial labels too. These hypothetical probabilities and budgets illustrate selection and accounting, not a measured improvement from active learning.

Illustrative case study

Fictional case: Annotation Bay uses uncertainty sampling for an inspection project. The system repeatedly chooses blurred images that experts cannot label consistently. The team initially interprets the large uncertainty as evidence that these are the most valuable examples.

It adds an annotation outcome for unusable inputs, reviews image quality, and compares query strategies that consider both uncertainty and coverage. A separate evaluation set remains unchanged throughout the rounds. No efficiency gain is claimed. The case illustrates that uncertain does not necessarily mean informative, and that the acquisition process must reflect the quality and cost of the labels it can obtain.

Common misconceptions and mistakes

Active learning is not the same as pseudo-labeling. In pseudo-labeling, a model supplies provisional labels; active learning selects examples for obtaining information from a label source.

High uncertainty does not automatically identify the most useful query. It can arise from noise, ambiguity, or inputs outside the intended task.

Selecting a batch by taking the top individual scores can produce redundant examples. Batch composition matters.

Finally, annotation count is not the only cost. Expert difficulty, review time, retraining, and repeated acquisition computations can change whether a strategy is worthwhile.

Business use

Deep active learning is worth investigating when labels are expensive and a substantial pool of relevant unlabeled examples is available. Define the annotation unit, quality standard, budget, and target evaluation metric before starting.

Compare with random sampling and practical stratified selection, including all initial labeling and repeated training costs. Track which groups or conditions are selected and which are neglected. Establish how ambiguous labels are resolved and when the loop stops. The operational objective is useful model improvement per available resource, not an impressive acquisition score detached from the actual task.

Check your understanding

A team reports needing only 50 queried labels but started with 500 labeled examples. What is missing from its annotation-budget claim?

Answer: The initial 500 labels. A fair comparison should account for the complete label budget, including initial data and any additional annotation or review costs.

Related terms

Active learning
acquisition function
uncertainty sampling
batch selection
core-set selection
annotation budget
model uncertainty
human-in-the-loop learning

Learn next

rank a small candidate pool by uncertainty, then inspect how many top-ranked examples are near-duplicates. Design a comparison with random selection using the same initial labels, cumulative budget, and untouched evaluation set.

Sources

Gal, Islam, and Ghahramani, Deep Bayesian Active Learning with Image Data: https://arxiv.org/abs/1703.02910 — primary abstract consulted for deep uncertainty-based acquisition.

Sener and Savarese, Active Learning for Convolutional Neural Networks: A Core-Set Approach: https://arxiv.org/abs/1708.00489 — primary abstract consulted for batch selection and representation coverage.

Supplied ontology context: https://raw.githubusercontent.com/berkeleybop/artificial-intelligence-ontology/main/aio.owl. This entry describes the learning workflow rather than treating it as one fixed network architecture. Examples and fictional case are original.

Previous: Deductive synthesis Next: Deep backward stochastic differential equation method