New LAD framework offers faithful, named explanations for neural networks

Researchers have introduced Language-Anchored Decomposition (LAD), a novel post-hoc framework designed to provide interpretable explanations for deep neural networks without altering the original model. LAD leverages large language models to generate concept vocabularies, which are then localized within image regions using CLIP-based similarity. By treating these language-grounded maps as a fixed constraint, LAD learns a concept basis that reconstructs the encoder's activations, ensuring that the derived concepts are both faithful to the model's behavior and human-interpretable. Experiments across various benchmarks demonstrate LAD's ability to produce spatially precise and decision-relevant explanations. AI

IMPACT Enables more trustworthy deployment of AI in high-stakes visual applications by providing interpretable and faithful explanations.

RANK_REASON The cluster contains a research paper detailing a new method for explaining deep neural network behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Ahsan Habib Akash
• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Hugging Face
• Language-Anchored Decomposition
• non-negative matrix factorization
• ScienceCast

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →