ReciproCAM: Lightweight Gradient-free Class Activation Map for Post-hoc Explanations

Seok-Yong Byun, Wonju Lee; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 8364-8370

To interpret model behavior AI practitioners have shed light on explainable AI techniques. While visual explanations like class activation maps (CAM) and its derivatives have demonstrated promise their applicability within post-hoc frameworks is often constrained by architectural limitations gradient computation capabilities or slow execution speeds. In this paper we propose a lightweight gradient-free ReciproCAM by spatially perturbing the internal feature map to exploit the correlation between activations and a model output. From the numerical results we achieve the gains of 1.78 to 3.72% in the ResNet family compared to ScoreCAM in average drop-coherence-complexity metric excluding the VGG-16 (1.39% drop) while ReciproCAM exhibits 148 times faster than ScoreCAM.