Researchers have developed SEAMS, a novel saliency method designed to identify image regions crucial for preserving a model's behavior. This approach optimizes a soft mask using a preservation objective, directly searching for compact masks that maintain specific model outputs like class probabilities or embeddings. SEAMS operates without needing auxiliary datasets or architecture-specific mechanisms, demonstrating its flexibility across different models such as ViT-S/16 and ConvNeXt. The method produces stable, interpretable, and competitive saliency maps, highlighting that visual explanations can be architecture-dependent. AI
IMPACT Provides a new method for understanding AI model decision-making, potentially improving interpretability and trust in computer vision systems.
RANK_REASON The cluster contains a research paper detailing a new method for AI model interpretability.
- alphaXiv
- arXiv
- CatalyzeX
- ConvNeXt
- DagsHub
- Gotit.pub
- Hugging Face
- Magdalena Trędowicz
- ScienceCast
- SEAMS
- ViT-S/16
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
