On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation

This paper tackles the problem of understanding and interpreting the classification decisions of automated image classification systems, which is valuable because it lets users verify a system's reasoning and gives human experts additional information. Since most machine-learning methods act as black boxes, the authors propose a general solution based on pixel-wise decomposition of nonlinear classifiers, introducing a methodology that visualizes the contributions of single pixels to predictions for both kernel-based classifiers over Bag-of-Words features and multilayered neural networks.

The per-pixel contributions are rendered as heatmaps that allow a human expert to intuitively verify the validity of a classification decision and to focus further analysis on regions of potential interest. The authors evaluate the method on classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten-digit dataset, and the pre-trained ImageNet model available in the Caffe open-source package, establishing a widely applicable approach to explaining classifier decisions.