Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift

Abstract

We might hope that when faced with unexpected inputs, well-designed software systems would fire off warnings. Machine learning (ML) systems, however, which depend strongly on properties of their inputs (e.g. the i.i.d. assumption), tend to fail silently. This paper explores the problem of building ML systems that fail loudly, investigating methods for detecting dataset shift, identifying exemplars that most typify the shift, and quantifying shift malignancy. We focus on several datasets and various perturbations to both covariates and label distributions with varying magnitudes and fractions of data affected. Interestingly, we show that across the dataset shifts that we explore, a two-sample-testing-based approach, using pre-trained classifiers for dimensionality reduction, performs best. Moreover, we demonstrate that domain-discriminating approaches tend to be helpful for characterizing shifts qualitatively and determining if they are harmful.

arXiv commentAdvances in Neural Information Processing Systems (NeurIPS) 2019

References in corpus (13)

• Scikit-learn: Machine Learning in Python
• Sequence to Sequence Learning with Neural Networks
• Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
• End to End Learning for Self-Driving Cars
• A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks
• Adversarial Attacks on Neural Networks for Graph Data
• Deep Anomaly Detection with Outlier Exposure
• On Causal and Anticausal Learning
• Uncertainty in the Variational Information Bottleneck
• Out-of-Distribution Detection using Multiple Semantic Label Representations
• Combining p-values via averaging
• A Less Biased Evaluation of Out-of-distribution Sample Detectors
• Classification accuracy as a proxy for two sample testing

Cited by in corpus (34)

• How I failed machine learning in medical imaging -- shortcomings and recommendations
• Prediction-Based Decisions and Fairness: A Catalogue of Choices, Assumptions, and Definitions
• Ethical Machine Learning in Health Care
• A survey on learning from imbalanced data streams: taxonomy, challenges, empirical study, and reproducible experimental framework
• Causality for Machine Learning
• Run-Time Monitoring of Machine Learning for Robotic Perception: A Survey of Emerging Trends
• GitTables: A Large-Scale Corpus of Relational Tables
• Diagnosis and Prognosis of COVID-19 Disease Using Routine Blood Values and LogNNet Neural Network
• Towards Robust and Reliable Algorithmic Recourse
• Domain Adaptation as a Problem of Inference on Graphical Models
• MultiBench: Multiscale Benchmarks for Multimodal Representation Learning
• CycleSense: Detecting Near Miss Incidents in Bicycle Traffic from Mobile Motion Sensors
• A unified framework for dataset shift diagnostics
• Adversarial Detection and Correction by Matching Prediction Distributions
• MMD Aggregated Two-Sample Test
• Newer is not always better: Rethinking transferability metrics, their peculiarities, stability and performance
• Chasing Your Long Tails: Differentially Private Prediction in Health Care Settings
• Algorithmic Recourse in the Wild: Understanding the Impact of Data and Model Shifts
• DeepLens: Interactive Out-of-distribution Data Detection in NLP Models
• Endogenous Macrodynamics in Algorithmic Recourse
• MLDemon: Deployment Monitoring for Machine Learning Systems
• Confidence-based Estimators for Predictive Performance in Model Monitoring
• InFlow: Robust outlier detection utilizing Normalizing Flows
• Sound-Dr: Reliable Sound Dataset and Baseline Artificial Intelligence System for Respiratory Illnesses
• Deploying clinical machine learning? Consider the following...
• High Dimensional Restrictive Federated Model Selection with multi-objective Bayesian Optimization over shifted distributions
• RapidRead: Global Deployment of State-of-the-art Radiology AI for a Large Veterinary Teleradiology Practice
• Tracking the risk of a deployed model and detecting harmful distribution shifts
• Failing Conceptually: Concept-Based Explanations of Dataset Shift
• The Influence of Faulty Labels in Data Sets on Human Pose Estimation
• Test for non-negligible adverse shifts
• Addressing the Between-Group Comparison Problem: Detecting Differences Between Correlation Matrix Populations due to Single-variable Perturbations for Resting State fMRI
• MAGDiff: Covariate Data Set Shift Detection via Activation Graphs of Deep Neural Networks
• Making Table Understanding Work in Practice