Reproduce “Mapping the Concept Landscape: Structural Perception of Global Distributions for Transparent Data Pruning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongyue Wu, Tao Ma
The Mapping the Concept Landscape (MCL) framework enables transparent data pruning for multimodal datasets by explicitly modeling fine-grained semantic concepts and their relationships. This approach preserves model performance using substantially reduced data, improving training efficiency and interpretability of data curation decisions.
Key insights
- Semantic diversity is compositional at a concept level, not adequately captured by single dense vector embeddings, necessitating explicit modeling of concepts for effective pruning.
- Formulating data pruning as a concept coverage problem with interpretable concept graphs offers transparency and allows direct optimization for semantic breadth.
- Prioritizing under-represented concepts and those involved in richer semantic interactions, while minimizing redundancy, leads to more informative and balanced pruned datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.