[Paper Review] Axiom-based Grad-CAM: Towards Accurate Visualization and Explanation of CNNs

Ruigang Fu, Qingyong Hu+4 more|arXiv (Cornell University)|Aug 5, 2020

Explainable Artificial Intelligence (XAI)13 references151 citations

TL;DR

XGrad-CAM introduces two axioms (Sensitivity and Conservation) to guide class-discriminative CNN visualizations, providing a theoretically grounded enhancement over Grad-CAM while remaining efficient and compatible with arbitrary CNNs.

ABSTRACT

To have a better understanding and usage of Convolution Neural Networks (CNNs), the visualization and interpretation of CNNs has attracted increasing attention in recent years. In particular, several Class Activation Mapping (CAM) methods have been proposed to discover the connection between CNN's decision and image regions. In spite of the reasonable visualization, lack of clear and sufficient theoretical support is the main limitation of these methods. In this paper, we introduce two axioms -- Conservation and Sensitivity -- to the visualization paradigm of the CAM methods. Meanwhile, a dedicated Axiom-based Grad-CAM (XGrad-CAM) is proposed to satisfy these axioms as much as possible. Experiments demonstrate that XGrad-CAM is an enhanced version of Grad-CAM in terms of conservation and sensitivity. It is able to achieve better visualization performance than Grad-CAM, while also be class-discriminative and easy-to-implement compared with Grad-CAM++ and Ablation-CAM. The code is available at https://github.com/Fu0511/XGrad-CAM.

Motivation & Objective

• Motivate the interpretability gap in CNN visualizations and establish two axioms (Sensitivity and Conservation) to guide CAM methods.
• Derive a linear CAM formulation that satisfies these axioms as much as possible.
• Propose XGrad-CAM (and Guided XGrad-CAM) as theoretically grounded, class-discriminative visualization tools for arbitrary CNNs.

Proposed method

• Formulate CAM as a linear combination of target-layer feature maps with weights w_c^k.
• Define weights by solving an optimization objective that enforces Sensitivity and Conservation (Eq. 4).
• Derive an approximate closed-form solution alpha_c^k (Eq. 7) to obtain XGrad-CAM (Eq. 8).
• Prove that XGrad-CAM generalizes CAM and reduces to CAM for GAP-CNNs (Appendix).
• Optional: combine XGrad-CAM with Guided Backprop to obtain Guided XGrad-CAM for finer details.

Experimental results

Research questions

• RQ1 Does XGrad-CAM satisfy the proposed axioms of Sensitivity and Conservation more effectively than existing CAM methods?
• RQ2 Is XGrad-CAM class-discriminative and efficient across common CNN architectures?
• RQ3 How does XGrad-CAM perform in localization (perturbation) analyses compared to Grad-CAM, Grad-CAM++, and Ablation-CAM?
• RQ4 What is the impact of the chosen target layer on visualization quality across architectures?

Key findings

• XGrad-CAM achieves better adherence to the two axioms (Sensitivity and Conservation) than Grad-CAM and Grad-CAM++ on the ILSVRC-12 validation set.
• In class-discrimination tests, Grad-CAM, Ablation-CAM, and XGrad-CAM show similar performance, while Grad-CAM++ performs worse.
• In perturbation analyses, XGrad-CAM yields higher confidence drop than Grad-CAM, with Ablation-CAM being more time-consuming; Grad-CAM++ may have higher raw performance but lower class-discriminability.
• XGrad-CAM provides stronger localization and is more efficient than Ablation-CAM (which is about 40x slower).
• Guided XGrad-CAM offers additional detail by combining XGrad-CAM with Guided Backprop.