[Paper Review] Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks

Haofan Wang, Zifan Wang+6 more|arXiv (Cornell University)|Oct 3, 2019

Explainable Artificial Intelligence (XAI)Computer Science38 references93 citations

TL;DR

Score-CAM introduces a gradient-free, score-based weighting of activation maps for post-hoc visual explanations, outperforming prior CAM/Grad-CAM variants in recognition and localization tasks.

ABSTRACT

Recently, increasing attention has been drawn to the internal mechanisms of convolutional neural networks, and the reason why the network makes specific decisions. In this paper, we develop a novel post-hoc visual explanation method called Score-CAM based on class activation mapping. Unlike previous class activation mapping based approaches, Score-CAM gets rid of the dependence on gradients by obtaining the weight of each activation map through its forward passing score on target class, the final result is obtained by a linear combination of weights and activation maps. We demonstrate that Score-CAM achieves better visual performance and fairness for interpreting the decision making process. Our approach outperforms previous methods on both recognition and localization tasks, it also passes the sanity check. We also indicate its application as debugging tools. Official code has been released.

Motivation & Objective

• Motivate better visual explanations for CNN decisions without gradient dependence.
• Propose a gradient-free weighting mechanism based on Increase of Confidence (CIC) per activation map.
• Show Score-CAM achieves superior fidelity and localization while passing sanity checks and enabling debugging tools.

Proposed method

• Define Increase of Confidence to quantify activation map importance.
• Compute CIC for each activation map by masking the input with the upsampled, normalized activation map and measuring class score change.
• Weight activation maps by softmax-normalized CIC scores and form the final map as a ReLU-weighted linear combination of activation maps.
• Use post-softmax scores as weights to improve class discrimination in the final explanation.
• Provide a practical algorithm (Algorithm 1) for implementing Score-CAM across arbitrary convolutional layers.
• Compare Score-CAM against Grad-CAM, Grad-CAM++, and other baselines on detection of object regions and localization.

Experimental results

Research questions

• RQ1 Does a gradient-free, score-based weighting of activation maps provide more faithful and discriminative visual explanations than gradient-based CAM variants?
• RQ2 How does Score-CAM perform on recognition accuracy, localization, and sanity checks compared to Grad-CAM/Grad-CAM++?
• RQ3 Can Score-CAM serve as a useful debugging tool for model misbehaviors and dataset biases?

Key findings

• Score-CAM yields smoother, less noisy saliency maps than several baselines.
• On recognition tasks, Score-CAM achieves higher Average Increase and lower Average Drop than competing methods (e.g., Average Drop 31.5% and Average Increase 30.6% for ScoreCAM vs. 63.5%/31.5% for others in the reported table).
• Score-CAM improves localization, with >60% energy of saliency maps falling inside target object bounding boxes on a subset of images.
• Score-CAM passes the sanity checks for saliency maps (sensitive to model parameters; passes randomization test).
• Score-CAM demonstrates stronger faithfulness in deletion/insertion evaluations (higher insertion AUC, lower deletion AUC compared to Grad-CAM/Grad-CAM++).
• Score-CAM can aid in debugging model misbehaviors and revealing dataset biases by comparing saliency maps across predicted vs. target classes.