Beyond Output Faithfulness: Learning Attributions that Preserve Computational Pathways

2025/09/04 by Siyu Zhang, Zhang, Siyu, Kenneth L. McMillan +1
Decision Sciences · Social Sciences · #Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Knowledge Management and Technology #Machine Learning (cs.LG) #Religion and Sociopolitical Dynamics in Nigeria

paper · pdf · doi:10.48550/arxiv.2509.04588

openalex publication_date 2025/09/04 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/30

Abstract

Faithfulness metrics such as insertion and deletion evaluate how feature removal affects model outputs but overlook whether explanations preserve the computational pathway the network actually uses. We show that external metrics can be maximized through alternative pathways -- perturbations that reroute computation via different feature detectors while preserving output behavior. To address this, we propose activation preservation as a tractable proxy for preserving computational pathways We introduce Faithfulness-guided Ensemble Interpretation (FEI), which jointly optimizes external faithfulness (via ensemble quantile optimization of insertion/deletion curves) and internal faithfulness (via selective gradient clipping). Across VGG and ResNet on ImageNet and CUB-200-2011, FEI achieves state-of-the-art insertion/deletion scores while maintaining significantly lower activation deviation, showing that both external and internal faithfulness are essential for reliable explanations.

Citations

• Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition
• Visual Explanations via Iterated Integrated Attributions
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations
• Governance Architecture for Neural Network Superposition: A Structural Solution to Hallucination via Routing and Interference Filtering
• Towards Better Explanations of Class Activation Mapping
• Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks
• Understanding Deep Networks via Extremal Perturbations and Smooth Masks
• Adversarial Examples Are Not Bugs, They Are Features
• Interpreting Adversarial Examples by Activation Promotion and Suppression
• Sanity Checks for Saliency Maps
• RISE: Randomized Input Sampling for Explanation of Black-box Models
• SmoothGrad: removing noise by adding noise
• A Unified Approach to Interpreting Model Predictions
• Network Dissection: Quantifying Interpretability of Deep Visual\n Representations
• Axiomatic Attribution for Deep Networks
• Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
• Learning Deep Features for Discriminative Localization
• Deep Residual Learning for Image Recognition
• Evaluating the visualization of what a Deep Neural Network has learned
• Striving for Simplicity: The All Convolutional Net
• Very Deep Convolutional Networks for Large-Scale Image Recognition
• ImageNet Large Scale Visual Recognition Challenge
• ImageNet Large Scale Visual Recognition Challenge
• Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
• Visualizing and Understanding Convolutional Networks

Related

• AttributionLab: Faithfulness of Feature Attribution Under Controllable Environments
• Incorporating Attribution Importance for Improving Faithfulness Metrics
• Normalized AOPC: Fixing Misleading Faithfulness Metrics for Feature Attribution Explainability
• RelP: Faithful and Efficient Circuit Discovery in Language Models via Relevance Patching
• Sampling Matters in Explanations: Towards Trustworthy Attribution Analysis Building Block in Visual Models through Maximizing Explanation Certainty