[Paper Review] IS-CAM: Integrated Score-CAM for axiomatic-based explanations

Rakshit Naidu, Ankita Ghosh+3 more|arXiv (Cornell University)|Oct 6, 2020

Explainable Artificial Intelligence (XAI)Computer Science12 references31 citations

TL;DR

IS-CAM integrates an axiomatic attribution concept within the Score-CAM pipeline to produce sharper, more faithful attribution maps across models, evaluated on ILSVRC 2012 validation images.

ABSTRACT

Convolutional Neural Networks have been known as black-box models as humans cannot interpret their inner functionalities. With an attempt to make CNNs more interpretable and trustworthy, we propose IS-CAM (Integrated Score-CAM), where we introduce the integration operation within the Score-CAM pipeline to achieve visually sharper attribution maps quantitatively. Our method is evaluated on 2000 randomly selected images from the ILSVRC 2012 Validation dataset, which proves the versatility of IS-CAM to account for different models and methods.

Motivation & Objective

• Motivate explainable AI for CNNs in sensitive applications.
• Develop a gradient-free attribution method with sharper localization.
• Integrate an axiomatic attribution concept within the Score-CAM framework.
• Quantitatively evaluate faithfulness and localization on ImageNet validation data.
• Demonstrate versatility across multiple CNN architectures.

Proposed method

• Introduce an integration operation inside the Score-CAM pipeline to compute averaged scores over multiple intervals of the input mask.
• Define N as the number of intervals in [0,1] and compute an integrated saliency map by averaging scores from normalized masks across intervals (M0=0).
• Use a normalization function s(A) = (A - min(A)) / (max(A) - min(A)) to focus on the object region.
• Produce the saliency map as a linear combination of average integrated scores and upsampled activation maps, followed by ReLU as in Score-CAM.
• Ground the approach in axiomatic attribution principles via integration over the input mask to obtain sharper attribution maps.
• Compare with Grad-CAM, Grad-CAM++, Smooth Grad-CAM++, Score-CAM, and SS-CAM on ImageNet validation images.

Experimental results

Research questions

• RQ1 Can integrating an axiomatic attribution approach within Score-CAM yield sharper and more faithful attribution maps?
• RQ2 How does IS-CAM perform in faithfulness and localization across multiple architectures on ImageNet data?
• RQ3 Does the integration interval parameter N significantly affect qualitative visuals or quantitative metrics?
• RQ4 How does IS-CAM compare to existing CAM-based methods in bounding-box localization tasks?

Key findings

• IS-CAM achieves competitive visual attribution quality and improves quantitative faithfulness versus several CAM baselines.
• In an Average Drop and Increase in Confidence evaluation on 2000 ImageNet images with VGG-16, IS-CAM shows strong performance relative to Score-CAM and SS-CAM.
• IS-CAM attains favorable AUC scores for insertion/deletion analyses compared with baseline CAM methods.
• Localization evaluation (Energy-based pointing game) shows IS-CAM achieving strong proportions across VGG-16, ResNet-18, and SqueezeNet variants, with the highest value on VGG-16 among listed methods.
• Overall, IS-CAM balances faithful saliency and localization across multiple architectures.