A new study published on arXiv investigates the impact of segmentation choices on the reliability of LIME explanations for document image classification. Researchers found that standard superpixel-based segmentations, commonly used for natural images, are poorly aligned with document structures like text regions and layout blocks. By comparing Quickshift and SLIC with document-aware segmentations derived from OCR bounding boxes and regular grids on the RVL-CDIP dataset, the study demonstrates that domain-specific segmentation significantly improves explanation consistency, correctness, and local fidelity. These document-aware methods also reveal shortcut behaviors related to dataset biases that are often missed by superpixel-based approaches, highlighting the need for interpretable representations tailored to the specific domain. AI
IMPACT Highlights the importance of domain-specific preprocessing for reliable AI model interpretability, potentially influencing how AI explanations are developed and evaluated.
RANK_REASON Academic paper on segmentation methods for AI model explanations. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
