Active Learning is the iterative machine learning framework where the model itself selects the most informative unlabeled examples to be annotated by a human oracle, minimizing the total labeling cost required to reach a target accuracy — transforming annotation from an exhaustive manual task into a targeted, model-guided process.

Why Random Labeling Is Wasteful

In a pool of 1 million unlabeled images, the vast majority are easy and redundant — the model already classifies them correctly with high confidence. Labeling those adds no new knowledge. Active learning identifies the critical minority of ambiguous, boundary-region examples where a human label provides the maximum information gain.

Core Query Strategies

• Uncertainty Sampling: Select the examples where the model is least confident. For classification, this means choosing the sample whose predicted class probability is closest to uniform (highest entropy). Simple, fast, and effective for many tasks.
• Query-by-Committee: Train an ensemble of models and select examples where the committee members disagree most. Disagreement signals that the training data does not yet constrain the hypothesis space in that region.
• Expected Model Change: Select the example that, if labeled and added to training, would cause the largest gradient update to the model parameters. Computationally expensive but directly targets informativeness rather than using uncertainty as a proxy.
• Diversity Sampling: Select a batch of examples that are both uncertain and diverse (spread across different regions of feature space), preventing the active learner from repeatedly querying a single ambiguous cluster.

The Active Learning Loop

1. Train the model on the current labeled set. 2. Apply the query strategy to rank all unlabeled examples. 3. Present the top-$k$ to the human annotator. 4. Add the newly labeled examples to the training set. 5. Retrain and repeat until the accuracy target is met or the annotation budget is exhausted.

Practical Pitfalls

• Cold Start: With very few initial labels, the model's uncertainty estimates are unreliable, causing poor initial selections. Warm-starting with a small random seed set (50-200 examples) is critical.
• Sampling Bias: Active learning selects a non-random subset of the data. Models trained on actively selected data may perform poorly on the true data distribution if the query strategy over-focuses on boundary cases.

Active Learning is the economically rational approach to annotation — replacing brute-force labeling budgets with intelligent, model-driven selection that achieves equivalent accuracy at 10-50% of the labeling cost.