The (Un)reliability of saliency methods

Abstract

Saliency methods aim to explain the predictions of deep neural networks. These methods lack reliability when the explanation is sensitive to factors that do not contribute to the model prediction. We use a simple and common pre-processing step ---adding a constant shift to the input data--- to show that a transformation with no effect on the model can cause numerous methods to incorrectly attribute. In order to guarantee reliability, we posit that methods should fulfill input invariance, the requirement that a saliency method mirror the sensitivity of the model with respect to transformations of the input. We show, through several examples, that saliency methods that do not satisfy input invariance result in misleading attribution.

Cited by in corpus (68)

• A Survey on Explainable Artificial Intelligence (XAI): Towards Medical XAI
• Pathologies of Neural Models Make Interpretations Difficult
• Explanations can be manipulated and geometry is to blame
• WT5?! Training Text-to-Text Models to Explain their Predictions
• Neural Network Attribution Methods for Problems in Geoscience: A Novel Synthetic Benchmark Dataset
• Acquisition of Chess Knowledge in AlphaZero
• Towards Automatic Concept-based Explanations
• Explaining Anomalies Detected by Autoencoders Using SHAP
• Adversarial Robustness as a Prior for Learned Representations
• Beyond saliency: understanding convolutional neural networks from saliency prediction on layer-wise relevance propagation
• Interpreting deep learning models for weak lensing
• Deep Weakly-Supervised Learning Methods for Classification and Localization in Histology Images: A Survey
• Stakeholders in Explainable AI
• Interpreting a Recurrent Neural Network's Predictions of ICU Mortality Risk
• Saccader: Improving Accuracy of Hard Attention Models for Vision
• Neural Network Attributions: A Causal Perspective
• Exploratory Not Explanatory: Counterfactual Analysis of Saliency Maps for Deep Reinforcement Learning
• What Do You See? Evaluation of Explainable Artificial Intelligence (XAI) Interpretability through Neural Backdoors
• Interpretability Beyond Classification Output: Semantic Bottleneck Networks
• Explaining Time Series Predictions with Dynamic Masks
• Smoothed Geometry for Robust Attribution
• Interpretable deep learning for nuclear deformation in heavy ion collisions
• What can AI do for me: Evaluating Machine Learning Interpretations in Cooperative Play
• Embedding Deep Networks into Visual Explanations
• SAUNet: Shape Attentive U-Net for Interpretable Medical Image Segmentation
• Global Explanations of Neural Networks: Mapping the Landscape of Predictions
• Interpretable Deep Learning under Fire
• Exact and Consistent Interpretation for Piecewise Linear Neural Networks: A Closed Form Solution
• On Explaining Your Explanations of BERT: An Empirical Study with Sequence Classification
• GCExplainer: Human-in-the-Loop Concept-based Explanations for Graph Neural Networks
• How to Manipulate CNNs to Make Them Lie: the GradCAM Case
• Best of both worlds: local and global explanations with human-understandable concepts
• Evolved Explainable Classifications for Lymph Node Metastases
• Towards Explainable Deep Learning for Credit Lending: A Case Study
• Discovering Molecular Functional Groups Using Graph Convolutional Neural Networks
• Finding and Visualizing Weaknesses of Deep Reinforcement Learning Agents
• Order in the Court: Explainable AI Methods Prone to Disagreement
• Robustness of Visual Explanations to Common Data Augmentation
• Semantics for Global and Local Interpretation of Deep Neural Networks
• Influence-Directed Explanations for Deep Convolutional Networks
• A Survey on the Robustness of Feature Importance and Counterfactual Explanations
• One Explanation is Not Enough: Structured Attention Graphs for Image Classification
• Identifying Galaxy Cluster Mergers with Deep Neural Networks using Idealized Compton-y and X-ray maps
• Interpreting Interpretations: Organizing Attribution Methods by Criteria
• Challenging common interpretability assumptions in feature attribution explanations
• Rethinking Positive Aggregation and Propagation of Gradients in Gradient-based Saliency Methods
• Gradient-based Analysis of NLP Models is Manipulable
• Towards Rigorous Interpretations: a Formalisation of Feature Attribution
• GANMEX: One-vs-One Attributions Guided by GAN-based Counterfactual Explanation Baselines
• YASENN: Explaining Neural Networks via Partitioning Activation Sequences
• The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?
• Regional Image Perturbation Reduces $L_p$ Norms of Adversarial Examples While Maintaining Model-to-model Transferability
• What Do Adversarially Robust Models Look At?
• ExAD: An Ensemble Approach for Explanation-based Adversarial Detection
• Understanding Regularization to Visualize Convolutional Neural Networks
• Cause and Effect: Hierarchical Concept-based Explanation of Neural Networks
• On Spectral Properties of Gradient-based Explanation Methods
• Heat and Blur: An Effective and Fast Defense Against Adversarial Examples
• DANCE: Enhancing saliency maps using decoys
• Harnessing value from data science in business: ensuring explainability and fairness of solutions
• Dependency Decomposition and a Reject Option for Explainable Models
• Training Machine Learning Models by Regularizing their Explanations
• An Overview of Computational Approaches for Interpretation Analysis
• Regression Concept Vectors for Bidirectional Explanations in Histopathology
• Semantic bottleneck for computer vision tasks
• Games for Fairness and Interpretability
• Corpus-level and Concept-based Explanations for Interpretable Document Classification
• Towards Explainable Fact Checking