Learning to Explain: An Information-Theoretic Perspective on Model Interpretation
arXiv:1802.07814
Abstract
We introduce instancewise feature selection as a methodology for model interpretation. Our method is based on learning a function to extract a subset of features that are most informative for each given example. This feature selector is trained to maximize the mutual information between selected features and the response variable, where the conditional distribution of the response variable given the input is the model to be explained. We develop an efficient variational approximation to the mutual information, and show the effectiveness of our method on a variety of synthetic and real data sets using both quantitative metrics and human evaluation.
arXiv commentAccepted to ICML 2018 as a long oral
References in corpus (12)
- Categorical Reparameterization with Gumbel-Softmax
- Striving for Simplicity: The All Convolutional Net
- Learning Important Features Through Propagating Activation Differences
- Convolutional Neural Networks for Sentence Classification
- The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
- Multiple Object Recognition with Visual Attention
- Online and Linear-Time Attention by Enforcing Monotonic Alignments
- A Hierarchical Neural Autoencoder for Paragraphs and Documents
- Investigating the influence of noise and distractors on the interpretation of neural networks
- Kernel Feature Selection via Conditional Covariance Minimization
- Variational Information Maximization for Feature Selection
- Greedy Attack and Gumbel Attack: Generating Adversarial Examples for Discrete Data
Cited by in corpus (18)
- Explainable Artificial Intelligence Approaches: A Survey
- Explainability in Music Recommender Systems
- F-BLEAU: Fast Black-box Leakage Estimation
- L-Shapley and C-Shapley: Efficient Model Interpretation for Structured Data
- Feature Removal Is a Unifying Principle for Model Explanation Methods
- Human Evaluation of Spoken vs. Visual Explanations for Open-Domain QA
- Neural Methods for Point-wise Dependency Estimation
- Data Representing Ground-Truth Explanations to Evaluate XAI Methods
- A Very Brief Introduction to Machine Learning With Applications to Communication Systems
- ASAC: Active Sensing using Actor-Critic models
- Neural Image Compression and Explanation
- DeepSmartFuzzer: Reward Guided Test Generation For Deep Learning
- Kernelized Hashcode Representations for Relation Extraction
- Bounded logit attention: Learning to explain image classifiers
- Learning Propagation Rules for Attribution Map Generation
- Get It Scored Using AutoSAS -- An Automated System for Scoring Short Answers
- Learning to Increase the Power of Conditional Randomization Tests
- Clusters in Explanation Space: Inferring disease subtypes from model explanations