Explainability methods show architecture-dependent performance across AI vision models

A new benchmark study published on arXiv investigates the effectiveness of explainable AI (XAI) attribution methods, particularly their transferability between Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). The research reveals that attribution performance is highly dependent on model architecture, and conclusions drawn from CNNs do not reliably apply to transformer-based models. The study highlights the need for architecture-aware, multi-dimensional evaluations, as conventional metrics may not accurately reflect localization or faithfulness across different model types. AI

IMPACT Findings challenge existing assumptions about AI model explainability, suggesting a need for architecture-specific evaluation methods.

RANK_REASON Academic paper presenting a controlled benchmark and new findings on AI model explainability. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Asanka G. Perera
• CNNS
• Hugging Face
• Vision Transformers

• paper
• other

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