Abstract:The transparent formulation of explanation methods is essential for elucidating the predictions of neural networks, which are typically black-box models. Layer-wise Relevance Propagation (LRP) is a well-established method that transparently traces the flow of a model's prediction backward through its architecture by backpropagating relevance scores. However, the conventional LRP does not fully consider the existence of skip connections, and thus its application to the widely used ResNet architecture has not been thoroughly explored. In this study, we extend LRP to ResNet models by introducing Relevance Splitting at points where the output from a skip connection converges with that from a residual block. Our formulation guarantees the conservation property throughout the process, thereby preserving the integrity of the generated explanations. To evaluate the effectiveness of our approach, we conduct experiments on ImageNet and the Caltech-UCSD Birds-200-2011 dataset. Our method achieves superior performance to that of baseline methods on standard evaluation metrics such as the Insertion-Deletion score while maintaining its conservation property. We will release our code for further research at <a class="link-external link-https" href="https://5ei74r0.github.io/lrp-for-resnet.page/" rel="external noopener nofollow">this https URL</a>

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: When the existing Layer - wise Relevance Propagation (LRP) method is applied to the ResNet architecture with skip connections, it fails to fully consider the existence of skip connections, resulting in unreliable interpretation results. Specifically, the traditional LRP method is mainly used for traditional convolutional neural networks (such as VGG), but performs poorly when dealing with complex architectures such as ResNet. This is mainly because the residual connections in ResNet create multiple non - linearly related paths, and these paths cannot be processed by the traditional relevance attribution process. To solve this problem, the author proposes the following improvement measures: 1. **Introduce Relevance Splitting**: A new relevance propagation rule - relevance splitting - is introduced at the place where the skip connection and the output of the residual block converge. This splitting ensures that the relevance maintains the conservation property throughout the propagation process, thereby ensuring the consistency and integrity of the interpretation. 2. **Ensure the Conservation Property**: The author emphasizes that the method they propose strictly adheres to the conservation property throughout the relevance propagation process, that is, the sum of the relevance received by each unit is equal to the sum of the relevance it redistributes. This ensures that the generated interpretation does not lose or add extra information. 3. **Heat Quantization**: In order to alleviate the problem of over - concentration in the irrelevant areas of the generated attribution map, the author introduces the heat quantization technique. By quantifying the attribution values, the attribution distribution becomes more uniform, thereby improving the quality of the attribution map. 4. **Evaluate the Effectiveness of the Method**: The author conducts experiments on the ImageNet and Caltech - UCSD Birds - 200 - 2011 datasets to verify the superior performance of the proposed method on standard evaluation metrics (such as Insertion - Deletion scores) and maintain the conservation property. Through these improvements, the author aims to provide a high - quality and transparent interpretation method for the ResNet model, thereby promoting the application of deep - learning models in key fields such as medical treatment and autonomous driving. ### Formula Summary 1. **Basic Formula of LRP**: \[ R_i^{(l)}=\sum_{j = 1}^E\frac{z_{ij}}{\sum_{k = 1}^D z_{kj}}R_j^{(l + 1)} \] where \(R_i^{(l)}\) and \(R_j^{(l+1)}\) represent the relevance scores of the \(l\)-th layer and the \((l + 1)\)-th layer respectively, and \(z_{ij}\) represents the contribution from \(h_i^{(l)}\) to \(h_j^{(l+1)}\). 2. **LRP Formula under Linear Projection**: \[ R_i^{(l)}=\sum_{j = 1}^E\frac{w_{ji}^+h_i^{(l)}}{\sum_{k = 1}^D w_{jk}^+h_k^{(l)}}R_j^{(l + 1)} \] where \(w_{ji}^+\) represents the non - negative part of the weight matrix \(W\). 3. **Relevance Splitting Formula**: - Symmetric Splitting: \[ (R_s)_i=(R_m)_i=\frac{R_i^{(l+1)}}{2} \] - Proportion - based Splitting: \[ (R_s)_i=R_i^{(l+1)}\cdot\frac{|(h_s)_i|}{|(h_m)_i|+|(h_s)_i|} \] \[ (R_m)_i=R_i^{(l+1)}\cdot\frac{|(h_m)_i|}{|(h_m)_i|+|(h_s)_i|} \] 4. **Heat Quantization Formula**: \[

Layer-Wise Relevance Propagation with Conservation Property for ResNet

LPNet: A Remote Sensing Scene Classification Method Based on Large Kernel Convolution and Parameter Fusion

Explaining Deep Learning Models for Structured Data using Layer-Wise Relevance Propagation

Neural network interpretability with layer-wise relevance propagation: novel techniques for neuron selection and visualization

Sparse Explanations of Neural Networks Using Pruned Layer-Wise Relevance Propagation

Advancing Attribution-Based Neural Network Explainability through Relative Absolute Magnitude Layer-Wise Relevance Propagation and Multi-Component Evaluation

AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers

Learning local discrete features in explainable-by-design convolutional neural networks

Fact or Artifact? Revise Layer-wise Relevance Propagation on various ANN Architectures

Mutual Information Preserving Back-propagation: Learn to Invert for Faithful Attribution

ExplainLFS: Explaining neural architectures for similarity learning from local perturbations in the latent feature space

ReX: A Framework for Generating Local Explanations to Recurrent Neural Networks

Prob-POS: A Framework for Improving Visual Explanations from Convolutional Neural Networks for Remote Sensing Image Classification

ParaLkResNet: an efficient multi-scale image classification network

Improving Network Interpretability via Explanation Consistency Evaluation

Residual Networks of Residual Networks: Multilevel Residual Networks

Enhancing the Extraction of Interpretable Information for Ischemic Stroke Imaging from Deep Neural Networks

Explaining deep residual networks predictions with symplectic adjoint method

Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning

LRNNet: A Light-Weighted Network with Efficient Reduced Non-Local Operation for Real-Time Semantic Segmentation

Towards explaining graph neural networks via preserving prediction ranking and structural dependency