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

Qinglong Zhang, Lu Rao+1 more|arXiv (Cornell University)|Mar 25, 2021

Explainable Artificial Intelligence (XAI)23 references33 citations

TL;DR

Group-CAM efficiently generates target-related saliency maps by grouping feature maps, de-noising initial masks, and weighting by class confidence after perturbations, achieving better explanations with fewer network queries than prior methods.

ABSTRACT

In this paper, we propose an efficient saliency map generation method, called Group score-weighted Class Activation Mapping (Group-CAM), which adopts the "split-transform-merge" strategy to generate saliency maps. Specifically, for an input image, the class activations are firstly split into groups. In each group, the sub-activations are summed and de-noised as an initial mask. After that, the initial masks are transformed with meaningful perturbations and then applied to preserve sub-pixels of the input (i.e., masked inputs), which are then fed into the network to calculate the confidence scores. Finally, the initial masks are weighted summed to form the final saliency map, where the weights are confidence scores produced by the masked inputs. Group-CAM is efficient yet effective, which only requires dozens of queries to the network while producing target-related saliency maps. As a result, Group-CAM can be served as an effective data augment trick for fine-tuning the networks. We comprehensively evaluate the performance of Group-CAM on common-used benchmarks, including deletion and insertion tests on ImageNet-1k, and pointing game tests on COCO2017. Extensive experimental results demonstrate that Group-CAM achieves better visual performance than the current state-of-the-art explanation approaches. The code is available at https://github.com/wofmanaf/Group-CAM.

Motivation & Objective

• Motivate efficient, discriminative visual explanations for CNN decisions to foster trust and interpretability.
• Develop a faster saliency method than RISE while maintaining or improving explanatory quality.
• Leverage a split-transform-merge strategy to produce group-based initial masks and weight them by class-confidence scores.

Proposed method

• Split class activations into G groups and sum gradients-weighted activations within each group to form initial masks.
• Apply de-noising to initial masks and normalize to create smooth, pixel-preserving masks.
• Blur unmasked regions of the input and compute confidence gains from perturbed inputs to weight masks.
• Aggregate masked inputs with weights (confidence scores) to form the final saliency map via a weighted sum.
• Compare efficiency and visual quality against RISE, XRAI, Grad-CAM, and Score-CAM on ImageNet-1k and COCO2017.

Experimental results

Research questions

• RQ1 Can Group-CAM produce more target-relevant saliency maps with significantly fewer network queries than existing methods?
• RQ2 Does the group-based masking plus perturbation approach improve both visual quality and localization performance compared to state-of-the-art explanations?
• RQ3 Can Group-CAM serve as an effective data augmentation technique for fine-tuning classifiers?

Key findings

• Group-CAM achieves higher overall explanatory performance (AUC) than Grad-CAM, Grad-CAM++, RISE, XRAI, and Score-CAM in insertion/deletion tests on ImageNet-1k, with Group-CAM reaching 56.8 insertion AUC, 12.3 deletion AUC, and 44.5 overall AUC.
• On ImageNet-1k, Group-CAM outperforms competitors in localization (pointing game) on COCO val2017, achieving 57.5 mean accuracy versus Grad-CAM’s 56.7.
• Group-CAM runs extremely fast (≈0.09 seconds per image with G=32) and is competitive with Grad-CAM in runtime, while delivering better explanations.
• Ablation studies show that growing the group count G improves overall scores, and a threshold theta around 70 yields strong performance.