Fibres of Failure: Classifying errors in predictive processes
arXiv:1803.00384
Abstract
We describe Fibres of Failure (FiFa), a method to classify failure modes of predictive processes using the Mapper algorithm from Topological Data Analysis. Our method uses Mapper to build a graph model of input data stratified by prediction error. Groupings found in high-error regions of the Mapper model then provide distinct failure modes of the predictive process. We demonstrate FiFa on misclassifications of MNIST images with added noise, and demonstrate two ways to use the failure mode classification: either to produce a correction layer that adjusts predictions by similarity to the failure modes; or to inspect members of the failure modes to illustrate and investigate what characterizes each failure mode.
arXiv comment10 pages, submitted to ICML 2018
References in corpus (27)
- Fast unfolding of communities in large networks
- Ensemble Adversarial Training: Attacks and Defenses
- Model-Agnostic Interpretability of Machine Learning
- "Why Should I Trust You?": Explaining the Predictions of Any Classifier
- Grad-CAM: Why did you say that?
- Learning with a Strong Adversary
- Adversarial Examples: Attacks and Defenses for Deep Learning
- Parseval Networks: Improving Robustness to Adversarial Examples
- Embedding Projector: Interactive Visualization and Interpretation of Embeddings
- An unexpected unity among methods for interpreting model predictions
- Investigating the influence of noise and distractors on the interpretation of neural networks
- Nothing Else Matters: Model-Agnostic Explanations By Identifying Prediction Invariance
- Interpretation of Prediction Models Using the Input Gradient
- Increasing the Interpretability of Recurrent Neural Networks Using Hidden Markov Models
- Using Visual Analytics to Interpret Predictive Machine Learning Models
- TreeView: Peeking into Deep Neural Networks Via Feature-Space Partitioning
- Interpretable Recurrent Neural Networks Using Sequential Sparse Recovery
- Interpreting the Predictions of Complex ML Models by Layer-wise Relevance Propagation
- Feature Importance Measure for Non-linear Learning Algorithms
- Interpretable Active Learning
- Visualizing Dynamics: from t-SNE to SEMI-MDPs
- Mapping chemical performance on molecular structures using locally interpretable explanations
- Interpreting extracted rules from ensemble of trees: Application to computer-aided diagnosis of breast MRI
- Toward Interpretable Topic Discovery via Anchored Correlation Explanation
- Visualizing textual models with in-text and word-as-pixel highlighting
- MDL-motivated compression of GLM ensembles increases interpretability and retains predictive power
- Interpretable Low-Dimensional Regression via Data-Adaptive Smoothing