SmoothGrad: removing noise by adding noise

About

Explaining the output of a deep network remains a challenge. In the case of an image classifier, one type of explanation is to identify pixels that strongly influence the final decision. A starting point for this strategy is the gradient of the class score function with respect to the input image. This gradient can be interpreted as a sensitivity map, and there are several techniques that elaborate on this basic idea. This paper makes two contributions: it introduces SmoothGrad, a simple method that can help visually sharpen gradient-based sensitivity maps, and it discusses lessons in the visualization of these maps. We publish the code for our experiments and a website with our results.

Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, Martin Wattenberg · 2017

Related benchmarks

Task	Dataset	Result
Image Classification	MNIST (test)	Accuracy90.68	894
Image Classification	SVHN (test)	Accuracy62.35	470
Explainability	ImageNet (val)	Insertion44.5	104
Localization	ImageNet-1k (val)	EHR—	79
Feature Relevance Evaluation	ImageNet (test)	R (Feature Relevance)0.38	60
Anomaly Detection	LAG (test)	Image-level AUROC—	57
Attribution Fidelity	ImageNet 1,000 images (val)	µFidelity0.23	48
Deletion	ImageNet 2,000 images (val)	Deletion Score0.13	48
Feature Attribution Evaluation	ImageNet standard (val)	AUC78.1	39
Data Attribution	CIFAR-10	AUC1.36	36