Seeing What Few-Shot Learners See: Contrastive Cross-Class Attribution for Explainability

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot learning (FSL) is vital for AI in data-limited domains like healthcare.
• Current FSL models lack transparency, hindering trust and deployment.
• Explainability methods are often architecture-specific and not broadly applicable.

Purpose of the Study:

• To introduce Contrastive Cross-Class Attribution (C3A) for enhanced explainability in FSL.
• To develop a classifier-agnostic approach applicable across various FSL architectures.
• To improve the interpretability and trustworthiness of FSL model predictions.

Main Methods:

• Proposed Contrastive Cross-Class Attribution (C3A) for query prediction explanation.
• Introduced C3A-E variant for episodic learning explanation.
• Developed contrastive local descriptors integrating NBNN and Fisher score for fine-grained details and inter-class discrimination.
• Ensured classifier-agnostic design leveraging backbone networks for broad applicability.

Main Results:

• Achieved significant improvements over state-of-the-art FSL explainability methods.
• Demonstrated absolute gains of 19.72% in insertion-area-under-curve (iAUC) and 25.88% in deletion-area-above-curve (dAAC).
• Validated C3A and C3A-E on four FSL benchmarks across diverse architectures (Conv64F, ResNet12, ResNet18, ViT).

Conclusions:

• C3A and C3A-E provide more stable and interpretable explanations than existing methods.
• The proposed methods offer deeper insights into the episodic learning process in FSL.
• The classifier-agnostic, posthoc design ensures seamless integration and broad applicability.