Mitigating Redundancy in Batch Active Learning

19 concepts16 papers1 proposal

Active learning aims to minimize data annotation costs by intelligently selecting the most informative samples from an unlabeled pool. In an ideal, purely sequential setting, this process is straightforward: a model identifies the single most uncertain or informative data point, receives a label from an oracle, updates its parameters, and repeats.

However, in modern machine learning workflows, sequential retraining is computationally prohibitive. Training deep neural networks or coordinating human annotation teams requires querying data in parallel batches.

This transition from sequential to batch selection introduces a critical failure mode: batch redundancy. If an acquisition function naively selects the top-$k$ most informative individual points based on the model's current state, it frequently selects a cluster of nearly identical samples. For instance, if a model is highly uncertain about a specific type of image artifact, a naive ranking strategy might fill an entire batch with minor variations of that exact same artifact. As noted by Azimi et al., this lack of diversity causes a massive drop in information efficiency compared to sequential methods, wasting a limited annotation budget on overlapping insights.

The Geometry of Information Overlap

Why does batch redundancy occur so consistently? The root cause lies in the geometry of the feature space. Highly informative points—such as those near a classifier's decision boundary or within regions of high data density—are rarely isolated. Instead, they cluster tightly together.

When an acquisition strategy evaluates points independently, it lacks global awareness of the batch's composition. As emphasized by Matsuo et al. and Zhao et al., selecting data purely on individual informativeness metrics pulls redundant samples from the same narrow region of the feature space. To maximize the utility of a batch, the selection mechanism must explicitly balance two competing forces: individual informativeness (how much a single point helps the model) and batch diversity (how distinct the selected points are from one another).

Naive Selection (High Informativeness, Low Diversity):
[Cluster of Ambiguous Points] -> Selects Point A1, A2, A3 (Redundant)

Diversity-Aware Selection (High Informativeness, High Diversity):
[Cluster A] -> Selects Point A1
[Cluster B] -> Selects Point B1
[Cluster C] -> Selects Point C1

Mathematical Frameworks for Joint Optimization

To resolve this tradeoff, researchers have developed several mathematical frameworks that move away from simple greedy ranking toward joint batch optimization.

Determinantal Point Processes (DPPs)

DPPs offer an elegant probabilistic framework to model diversity. A DPP defines a probability distribution over subsets of data where the probability of selecting a specific subset is proportional to the determinant of a similarity kernel matrix. The diagonal elements of this matrix capture individual informativeness, while the off-diagonal elements capture pairwise similarity.

When computing the determinant, highly correlated points collapse the volume of the spanned space, severely penalizing the subset's overall score. Kim et al. utilized DPPs over a multidimensional meta-information space to prevent ranking methods from collapsing into clusters of similar difficult samples. Similarly, Biyik et al. leveraged DPPs to parallelize queries, ensuring that the system does not waste human attention on overlapping questions.

Information-Theoretic and Matrix Formulations

Another approach leverages mutual information and joint entropy. Instead of evaluating points independently, these methods optimize the collective information gain of the entire batch. For example, Yuhong Guo framed batch selection as a matrix partition problem within a Gaussian Process framework to maximize mutual information between labeled and unlabeled sets without needing to know the downstream classifier.

In the context of Bayesian optimization, Li et al. demonstrated that information-theoretic acquisition functions like entropy search inherently penalize highly correlated points, ensuring that a batch spans diverse inputs or fidelity levels. Aronsson et al. similarly applied expected information gain via mean-field approximations to maintain diversity in active correlation clustering.

Continuous and Dual Optimization

Rather than relying on discrete combinatorial searches, batch selection can be cast as continuous optimization. Guo et al. formulated batch selection as a continuous optimization problem that directly minimizes a bound on the classification error while accounting for information overlap.

More recently, Elenter et al. introduced a framework that treats data samples as constraints in an optimization problem, using Lagrangian dual variables as a direct proxy for informativeness to balance individual utility with batch-wide diversity. From a statistical perspective, Gentile et al. utilized G-optimal design within a stage-wise greedy algorithm to achieve theoretically optimal minimax convergence rates, proving that formal diversity guarantees prevent statistical redundancy even in the presence of noisy labels.

Broad Applicability Across Domains

The necessity of balancing informativeness and diversity spans diverse machine learning paradigms, demonstrating that batch redundancy is a fundamental geometric challenge rather than a domain-specific quirk.

  • Preference Learning and Robotics: When teaching robots human reward functions via pairwise comparisons, asking nearly identical questions frustrates users and wastes human effort. Biyik et al. and Talati et al. showed that maximizing the distance between high-entropy queries ensures each question provides unique insights into the underlying reward function.
  • Computer Vision and Medical Imaging: In histopathological image classification, Cooper et al. demonstrated that decomposing uncertainty prevents standard active learning algorithms from getting distracted by redundant, ambiguous artifacts, ensuring a broader coverage of the data distribution.
  • Natural Language Processing: For complex tasks like multi-label text classification, Tan et al. noted that relying purely on raw model features for diversity can be noisy, requiring robust scoring rules to ensure the batch covers the entire label space effectively.

By shifting the focus from independent ranking to collective utility, these diverse methodologies ensure that each sample in a batch provides a unique, non-overlapping contribution to the model's evolving understanding of the data distribution.

Key Foundational Work

Discriminative Batch Mode Active Learning (2007) — contributed 1 concept elements to this direction.
Active Instance Sampling via Matrix Partition (2010) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_diversity_informativeness_tradeoff_329bb76f,
  title = {diversity-informativeness tradeoff},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/diversity-informativeness-tradeoff/rdr_329bb76fcc6c42b68481d8a7ee53ea36},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_diversity_informativeness_tradeoff_329bb76f_v1,
  title = {diversity-informativeness tradeoff},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/diversity-informativeness-tradeoff/rdr_329bb76fcc6c42b68481d8a7ee53ea36/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake