feiyang-k/curation-bench/active-learning-for-convolutional-neural-networks-a-core-set-approach-arxiv-1708-00489v4

Active Learning for Convolutional Neural Networks: A Core-Set Approach

One-line decision

Use this skill when you want to select the most informative samples for annotation using coreset-based active learning in embedding space. Avoid it when you have unlimited annotation budget or random selection is sufficient.

Skill metadata

• Skill type: coreset-active-learning
• Paper kind: operational-method
• Actionability: high
• Evidence quality: full_paper

Goal

Select the most informative samples for annotation using a core-set approach that maximizes coverage of the embedding space, enabling efficient training with fewer labeled examples.

Problem signature

• Modality: images or any data with embeddings; applicable to VLM data.
• Data state: large unlabeled pool with budget for selective annotation.
• Scale regime: thousands to millions of candidate samples.
• Model requirement: Any model with embedding extraction capability.

Use when

• You have a limited annotation budget.
• You want to maximize information gain from each annotated sample.
• You can compute embeddings for your data pool.

Do not use when

• You have unlimited annotation budget.
• Random selection is sufficient for your needs.
• You cannot compute embeddings.

Required inputs

• unlabeled_pool: Large pool of unlabeled data.
• embedding_model: Model for computing data embeddings.
• annotation_budget: Number of samples to select for annotation.

Optional inputs

• initial_labeled_set: Small initially labeled set.

Outputs

• selected_samples: Samples selected for annotation.
• coreset: Representative subset covering the embedding space.

Assumptions and prerequisites

• Core-set selection in embedding space maximizes data coverage.
• Coverage correlates with training utility.
• Embeddings capture relevant data characteristics.

Procedure

1. Compute embeddings

Action: Embed all candidate samples using a pre-trained model. Why: Embeddings enable similarity computation. Note: See paper for details.

1. Solve core-set problem

Action: Select samples that minimize the maximum distance from any point to the nearest selected point. Why: Core-set maximizes coverage of the data distribution. Note: See paper for details.

1. Annotate selected samples

Action: Label the selected samples. Why: Annotation provides training supervision. Note: See paper for details.

1. Train and iterate

Action: Train on annotated data and repeat selection. Why: Iterative selection adapts to model improvements. Note: See paper for details.

Parameters to set

• budget_per_round — Role: Samples to select per active learning round. How to set: Based on annotation capacity. Default/range: Task-dependent. Effect: Larger budgets provide more data per round.
• embedding_model — Role: Model for computing embeddings. How to set: Use the best available pre-trained model. Default/range: Task-dependent. Effect: Better embeddings improve selection quality.

Validation checks

• Core-set selection should outperform random selection.
• Coverage should increase with each selection round.
• Training performance should improve efficiently.

Failure modes

• Embedding quality limits selection quality.
• The core-set problem is NP-hard; approximations may be suboptimal.
• Outliers may be over-selected.

Adaptation notes for VLM training

• Apply core-set active learning to VLM data annotation.
• Use CLIP embeddings for multimodal core-set selection.
• Combine with other selection criteria for hybrid strategies.

Implementation notes

• Use efficient k-center algorithms for core-set selection.
• Cache embeddings for repeated selection rounds.
• Track coverage metrics during selection.

Evidence from the paper

• Core-set active learning outperforms random and uncertainty-based selection.
• The approach maximizes coverage of the data distribution.
• Embedding-based selection enables efficient annotation.
• The method scales to large unlabeled pools.

Source paper

• Title: Active Learning for Convolutional Neural Networks: A Core-Set Approach
• Year: 2018
• Venue: ICLR
• Paper ID: arxiv-1708.00489v4
• URL: http://arxiv.org/abs/1708.00489v4
• arXiv ID: 1708.00489v4