CA-Stream: Attention-based pooling for interpretable image recognition

Felipe Torres (1) , Hanwei Zhang (1) , Ronan Sicre (1) , Stéphane Ayache (1) , Yannis Avrithis (2)

1 QARMA - éQuipe d'AppRentissage de MArseille
2 IARAI - Institute of Advanced Research in Artificial Intelligence [Vienna]

Explanations obtained from transformer-based architectures in the form of raw attention, can be seen as a class-agnostic saliency map. Additionally, attention-based pooling serves as a form of masking the in feature space. Motivated by this observation, we design an attention-based pooling mechanism intended to replace Global Average Pooling (GAP) at inference. This mechanism, called Cross-Attention Stream (CA-Stream), comprises a stream of cross attention blocks interacting with features at different network depths. CA-Stream enhances interpretability in models, while preserving recognition performance.

Mots clés: EXplainable AI XAI; interpretability; attention-based models; image classification

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Felipe Torres, Hanwei Zhang, Ronan Sicre, Stéphane Ayache, Yannis Avrithis. CA-Stream: Attention-based pooling for interpretable image recognition. XAI4CV workshop (CVPR), Jun 2024, Seatle, WA, United States. ⟨hal-04551613⟩