ResearchNovember 17, 2026
11
🌱
Attribution Fails at the Decoder
Attributing through the decoder yields sharper explanations for structured outputs.
#attribution#interpretability#computer vision#Grad-CAM

🔥 What happened
Zaer Bin Zaqy and Brinnae Bent show in a new arXiv paper that attribution methods for structured outputs like gaze or rotation predict raw coordinates, not the final prediction. Their decoder-aware tangent Grad-CAM yields 44% greater prediction displacement under 10% pixel deletion on MPIIGaze.
💡 Why it matters
Current methods dump up to 72% of relevance mass onto irrelevant regions—pure waste. The fix: attribute through the decoder differential, which for normalization reduces to a scaled tangent projection. On PASCAL3D+, this improves mean deletion displacement by 14.1 degrees.
âš¡ Our take
If you build explainability for 3D vision, stop explaining raw coordinates. The decoder isn't a detail—it's half the truth.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.