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

Analysis

Why This Paper Matters

This paper addresses a critical gap in machine learning: the lack of interpretability in nonlinear classifiers, particularly deep neural networks. As AI systems increasingly impact high-stakes domains like medical imaging and autonomous driving, understanding why a model makes a specific decision is as important as the decision itself. Prior to this work, most explanation methods were limited to linear models or simple feature importance measures. By introducing Layer-Wise Relevance Propagation (LRP), the authors provided a principled, theoretically grounded approach to decompose any nonlinear classifier's output into pixel-level contributions.

The significance is amplified by the paper's timing (2015), when deep learning was rapidly gaining traction but interpretability tools were scarce. LRP offered a practical solution that could be applied to state-of-the-art models like the Caffe ImageNet network, making it immediately useful to practitioners. The method's generality—covering both kernel methods and neural networks—ensured broad applicability across the field.

Technical Contributions

• Layer-Wise Relevance Propagation (LRP): A backward propagation algorithm that redistributes the prediction score from the output layer to the input layer using conservation-based rules. Each neuron's relevance is computed as a weighted sum of higher-layer relevances, ensuring the total relevance is preserved.
• Handling of Nonlinearities: LRP handles ReLU, max-pooling, and other nonlinear operations by using local redistribution rules that respect the network's structure, such as the epsilon rule for stabilizing divisions.
• Extension to Kernel Classifiers: For Bag of Words models with RBF kernels, LRP decomposes the kernel evaluation into contributions from individual features, enabling pixel-wise heatmaps.
• Visualization as Heatmaps: The output is a heatmap where positive (red) and negative (blue) relevance scores indicate pixels that support or oppose the classification decision, respectively.

Results

The paper evaluates LRP on four datasets: synthetic geometric shapes (where ground truth is known), MNIST handwritten digits, PASCAL VOC 2009 object detection, and the ImageNet ILSVRC-2012 challenge using the Caffe reference model. On synthetic data, LRP heatmaps correctly highlight the shape's interior and edges. For MNIST, the method focuses on the digit's stroke regions. On PASCAL VOC, LRP identifies object parts (e.g., wheels of a car) that drive classification. For ImageNet, the heatmaps align with human-annotated object bounding boxes, demonstrating practical utility. No quantitative metrics like accuracy or IoU are reported; evaluation is qualitative via visual inspection.

Significance

LRP has become a cornerstone of explainable AI, inspiring numerous follow-up works such as DeepLIFT, Integrated Gradients, and Grad-CAM. Its conservation principle influenced the theoretical understanding of attribution methods. The paper's clear exposition and open-source implementation accelerated adoption in fields like medical imaging, where interpretability is mandatory. Today, LRP remains a standard baseline for evaluating new explanation techniques and is integrated into popular frameworks like iNNvestigate and Captum.