ExplainabilityTransparencyAccountabilitycs.CV
LaFAM: Unsupervised Feature Attribution with Label-free Activation Maps
Aray Karjauv, Sahin Albayrak
July 8, 2024
Deutsche Jahrestagung für Künstliche Intelligenz
4 citations
Key Takeaways
-
LaFAM offers a simpler path to explainability: The method streamlines existing Class Activation Map (CAM) approaches by using raw activation maps directly, eliminating the need for labeled data—making it more practical for real-world applications where labels may be unavailable or expensive to obtain.
-
Particularly valuable for self-supervised learning models: The paper demonstrates that LaFAM is especially effective at generating saliency maps (visual explanations of what a model focuses on) for self-supervised learning systems, which are increasingly used in industry but have historically been harder to interpret.
-
Maintains performance across different learning paradigms: LaFAM works in both supervised and self-supervised learning contexts without sacrificing effectiveness, suggesting it could be a broadly applicable tool for understanding what different types of AI models are actually "looking at" when making decisions.
-
Addresses an underexplored gap in explainability research: The authors note that while activation maps are known to reveal important features in deep learning models, their direct use for feature attribution has been overlooked—this work fills that gap with a more efficient approach than conventional methods.
Abstract
Convolutional Neural Networks (CNNs) are known for their ability to learn hierarchical structures, naturally developing detectors for objects, and semantic concepts within their deeper layers. Activation maps (AMs) reveal these saliency regions, which are crucial for many Explainable AI (XAI) methods. However, the direct exploitation of raw AMs in CNNs for feature attribution remains underexplored in literature. This work revises Class Activation Map (CAM) methods by introducing the Label-free Activation Map (LaFAM), a streamlined approach utilizing raw AMs for feature attribution without reliance on labels. LaFAM presents an efficient alternative to conventional CAM methods, demonstrating particular effectiveness in saliency map generation for self-supervised learning while maintaining applicability in supervised learning scenarios.