Explaining Image Classifiers by Counterfactual Generation

Abstract

When an image classifier makes a prediction, which parts of the image are relevant and why? We can rephrase this question to ask: which parts of the image, if they were not seen by the classifier, would most change its decision? Producing an answer requires marginalizing over images that could have been seen but weren't. We can sample plausible image in-fills by conditioning a generative model on the rest of the image. We then optimize to find the image regions that most change the classifier's decision after in-fill. Our approach contrasts with ad-hoc in-filling approaches, such as blurring or injecting noise, which generate inputs far from the data distribution, and ignore informative relationships between different parts of the image. Our method produces more compact and relevant saliency maps, with fewer artifacts compared to previous methods.

arXiv commentICLR 2019 Camera Ready

Cited by in corpus (32)

• Explaining by Removing: A Unified Framework for Model Explanation
• DeepAID: Interpreting and Improving Deep Learning-based Anomaly Detection in Security Applications
• Explaining Classifiers with Causal Concept Effect (CaCE)
• Counterfactual State Explanations for Reinforcement Learning Agents via Generative Deep Learning
• On the (In)fidelity and Sensitivity for Explanations
• Getting a CLUE: A Method for Explaining Uncertainty Estimates
• Counterfactual Explanation Based on Gradual Construction for Deep Networks
• Generative causal explanations of black-box classifiers
• Explanation by Progressive Exaggeration
• Shapley explainability on the data manifold
• Explaining Visual Models by Causal Attribution
• ECINN: Efficient Counterfactuals from Invertible Neural Networks
• What went wrong and when? Instance-wise Feature Importance for Time-series Models
• Trustworthy Convolutional Neural Networks: A Gradient Penalized-based Approach
• Proactive Pseudo-Intervention: Causally Informed Contrastive Learning For Interpretable Vision Models
• Human-interpretable model explainability on high-dimensional data
• Explanations Based on Item Response Theory (eXirt): A Model-Specific Method to Explain Tree-Ensemble Model in Trust Perspective
• Dynamic Measurement Scheduling for Event Forecasting using Deep RL
• Robust Models Are More Interpretable Because Attributions Look Normal
• The Out-of-Distribution Problem in Explainability and Search Methods for Feature Importance Explanations
• Resisting Out-of-Distribution Data Problem in Perturbation of XAI
• In-Distribution Interpretability for Challenging Modalities
• Contextual Prediction Difference Analysis for Explaining Individual Image Classifications
• On Quantitative Evaluations of Counterfactuals
• Born Identity Network: Multi-way Counterfactual Map Generation to Explain a Classifier's Decision
• Designing Counterfactual Generators using Deep Model Inversion
• Investigating and Simplifying Masking-based Saliency Methods for Model Interpretability
• Towards Robust Classification Model by Counterfactual and Invariant Data Generation
• Neural-to-Tree Policy Distillation with Policy Improvement Criterion
• DANCE: Enhancing saliency maps using decoys
• Explaining the Road Not Taken
• Detecting and Understanding Generalization Barriers for Neural Machine Translation