SUNY: A Visual Interpretation Framework for Convolutional Neural Networks from a Necessary and Sufficient Perspective

Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, Zhaodan Kong, Kwan-Liu Ma; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 8371-8376

In spite of the ongoing evolution of deep learning Convolutional Neural Networks (CNNs) remain the de facto choice for numerous vision applications. To foster trust researchers have proposed various methods for visually interpreting CNNs via heatmaps which highlight the input regions important to a specific model decision. However in terms of the underlying design logic existing approaches often concentrate on model parameters overlooking the fundamental "why" question integral to human cognition. Thus they fail to embrace the two critical and complementary sides in reasoning: necessity and sufficiency. To address these issues we introduce SUNY a framework designed to rationalize the explanations toward better human understanding from both necessary and sufficient perspectives in a bi-directional manner. Extensive evaluations justify that SUNY not only yields more informative and convincing explanations from both angles but also achieves performances competitive to other approaches across different CNN architectures over different datasets.