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

TL;DR

This work tackles explainability in deep neural networks by introducing absLRP, a layer-wise relevance propagation rule that accounts for absolute activation magnitudes to produce contrastive, sparse attributions across CNNs and transformers. It also proposes Global Attribution Evaluation (GAE), a single-score metric that fuses Local Consistency (robustness and faithfulness via MoRF/LeRF) and Contrastiveness into $GAE(m_A)=LC(m_A)\cdot C(m_A)$, enabling fair cross-model comparisons. Through extensive experiments on ImageNet and PascalVOC with VGG, ResNet50, and ViT-Base, absLRP often achieves top performance on standard metrics and substantially improves attribution quality, including pixel-level, noise-free maps for Vision Transformers. Ablation and qualitative analyses reveal the necessity of propagating through all Transformer components and demonstrate the method’s broad applicability, including potential extension to text and other modalities.

Abstract

Recent advancement in deep-neural network performance led to the development of new state-of-the-art approaches in numerous areas. However, the black-box nature of neural networks often prohibits their use in areas where model explainability and model transparency are crucial. Over the years, researchers proposed many algorithms to aid neural network understanding and provide additional information to the human expert. One of the most popular methods being Layer-Wise Relevance Propagation (LRP). This method assigns local relevance based on the pixel-wise decomposition of nonlinear classifiers. With the rise of attribution method research, there has emerged a pressing need to assess and evaluate their performance. Numerous metrics have been proposed, each assessing an individual property of attribution methods such as faithfulness, robustness or localization. Unfortunately, no single metric is deemed optimal for every case, and researchers often use several metrics to test the quality of the attribution maps. In this work, we address the shortcomings of the current LRP formulations and introduce a novel method for determining the relevance of input neurons through layer-wise relevance propagation. Furthermore, we apply this approach to the recently developed Vision Transformer architecture and evaluate its performance against existing methods on two image classification datasets, namely ImageNet and PascalVOC. Our results clearly demonstrate the advantage of our proposed method. Furthermore, we discuss the insufficiencies of current evaluation metrics for attribution-based explainability and propose a new evaluation metric that combines the notions of faithfulness, robustness and contrastiveness. We utilize this new metric to evaluate the performance of various attribution-based methods. Our code is available at: https://github.com/davor10105/relative-absolute-magnitude-propagation

Figures (16)

• Figure 1: Toy examples illustrating key differences between LRP-$\alpha_1\beta_0$ and our method. White circles represent inputs and layer activations, with weights on connecting lines, while final attribution values are shown in yellow circles. In the left example, equal relative absolute magnitudes result in similar attributions for LRP-$\alpha_1\beta_0$ and our method. In the right example, vastly different activation magnitudes lead our method to consider the larger absolute activation of the bottom hidden neuron in the attribution map.
• Figure 2: Change curves for an example from ImageNet - VGG - absLRP: (a) MoRF and LeRF change curves for model output (b) MoRF and LeRF change curves for attribution maps (c) Difference curves for model output and attribution maps. Smaller difference in these curves signals a more locally robust method.
• Figure 3: Impact maps on an example from ImageNet - VGG network: (a) Original image (b) Sum of per step absolute products of input and gradient when masking low impact features first (c) Masking high impact features first (d) Difference between (a) and (b) (e) Sign of (d)
• Figure 4: (a) Example mosaic image - ImageNet (b) Our method's attribution map (positive only) - selected bottom left image as target class (c) LRP-$\alpha_2\beta_1$ attribution map (positive only), example of an attribution method with poor contrastive properties
• Figure 5: absLRP attribution maps for VGG - ImageNet