Reproduce “Enhancing Performance of Explainable AI Models with Constrained Concept Refinement” 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

Geyu Liang, Senne Michielssen, Salar Fattahi

This paper introduces Constrained Concept Refinement (CCR), a framework designed to optimize concept embeddings in interpretable-by-design AI models, improving predictive accuracy while preserving interpretability. CCR achieves notable accuracy gains and up to a tenfold reduction in computational cost on image classification benchmarks, supported by theoretical guarantees for both performance improvement and concept alignment.

Key insights

  • Small deviations in observed concept representations from ground truth latent features lead to a quantifiable and proportional degradation in model performance, highlighting the necessity of concept refinement.
  • The proposed Constrained Concept Refinement (CCR) framework can simultaneously enhance prediction accuracy and improve the faithfulness of concept interpretations by iteratively refining concept embeddings under interpretability-preserving constraints.
  • Leveraging multiple input samples allows CCR to achieve geometric convergence of refined concepts towards ground truth and linear convergence of prediction loss to zero, enabling both high accuracy and clear explanations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.