Active learning operates on a simple, compelling premise: instead of labeling an entire massive dataset, a machine learning model should selectively query only the most informative data points. The most intuitive metric for informativeness is uncertainty. If a model is highly unsure about a sample's classification—perhaps because it lies directly on the current decision boundary—labeling that sample should provide the maximum possible information gain.
This logic holds perfectly when querying data points one by one. However, in practical deep learning workflows, retraining a model after every single new label is computationally prohibitive. Instead, practitioners use batch-mode active learning, querying a batch of $B$ samples simultaneously to maximize throughput.
This is where pure uncertainty sampling breaks down. If a model is highly uncertain about a specific data point, it will be equally uncertain about its near-identical neighbors in the feature space. As noted by Sudhanshu Mittal et al., an algorithm that simply selects the top $B$ most uncertain points will often suffer from "mode collapse," picking a tight cluster of highly similar samples. For instance, in a video dataset, an uncertainty-only strategy might select twenty consecutive frames of the same action, wasting the annotation budget on redundant information that fails to broaden the model's understanding.
The Mechanism of Batch Redundancy
To understand why this happens, consider the geometry of the feature space. High-uncertainty regions are often concentrated around specific, complex decision boundaries or noisy outliers. When an acquisition function evaluates samples independently, multiple data points from the same dense, unresolved region will all receive high uncertainty scores.
As Yifan Fu et al. demonstrate, this independent ranking causes the selected batch to cluster tightly together. This phenomenon creates several distinct failure modes across different domains:
- Information Overlap: In physiological monitoring, Zeeshan Syed observed that purely uncertainty-driven queries result in significant information overlap, where multiple samples provide the same underlying signal.
- Wasted Annotation Budget: In video analysis, such as action spotting in football videos, Silvio Giancola et al. and Debanjan Goswami point out that selecting highly similar frames does not help the model generalize across the full data distribution.
- Narrow Specialization: For large language models (LLMs), Arnav Das shows that selecting data based solely on model uncertainty results in a redundant training set that lacks task diversity, focusing too narrowly on a small set of confusing examples.
Ultimately, while each sample in an uncertainty-selected batch might be highly informative individually, they are highly redundant as a group.
The Uncertainty-Diversity Trade-off
To overcome batch redundancy, active learning must balance two competing dimensions:
- Informativeness (Uncertainty): Targeting regions where the model's current predictions have high entropy or variance.
- Representativeness (Diversity): Ensuring that the selected batch spans the broader data distribution and covers unexplored regions of the feature space.
This is known as the uncertainty-diversity trade-off. As argued by Kwanyoung Kim et al., focusing exclusively on distribution representativeness can cause the model to ignore task-specific difficulties, while focusing solely on uncertainty leads to redundant outliers. A successful batch selection strategy must treat the batch as a collective unit rather than a collection of independent individuals.
Algorithmic Frameworks for Balancing the Trade-off
Researchers have developed several elegant mechanisms to mathematically enforce diversity while maintaining a focus on uncertainty.
Submodular Maximization
One of the most mathematically rigorous ways to handle redundancy is through submodularity, a property often described as the law of diminishing returns. Kai Wei et al. introduced the FASS (Filtered Active Submodular Selection) framework, which first filters the unlabeled pool to retain high-uncertainty samples and then applies submodular maximization to select a representative subset. Because submodular functions explicitly penalize the addition of elements that are similar to already selected elements, this naturally prevents batch redundancy.
Pre-Clustering and Hybrid Metrics
Another intuitive approach is to structure the feature space before evaluating uncertainty. Zhixin Huang et al. proposed the UUPC framework, which uses unsupervised clustering to partition the data space. By selecting uncertain samples across different clusters, the algorithm ensures that the queried batch is distributed globally rather than concentrated locally. Similarly, Ehsan Elhamifar et al. formulated batch selection as a convex optimization problem that incorporates a dissimilarity measure alongside confidence scores, ensuring the chosen batch covers unexplored regions.
Stochastic and Density-Aware Sampling
Instead of optimizing complex discrete functions, some methods introduce diversity through stochasticity or density awareness. Herve Lombaert introduced Stochastic Batches, a method that evaluates random candidate batches and selects the one that maximizes collective uncertainty. This simple randomization naturally breaks up tight clusters of redundant samples (such as consecutive slices in a 3D MRI scan). From a geometric perspective, density-aware core-set approaches leverage the insight that samples residing in sparse regions of the feature space are inherently less redundant and more informative, providing a geometric solution to the mode collapse problem.
By explicitly modeling the relationship between samples in a batch, these modern active learning frameworks ensure that every single label paid for contributes entirely new insight to the model.