WILDS: A Benchmark of in-the-Wild Distribution Shifts

Abstract

Distribution shifts -- where the training distribution differs from the test distribution -- can substantially degrade the accuracy of machine learning (ML) systems deployed in the wild. Despite their ubiquity in the real-world deployments, these distribution shifts are under-represented in the datasets widely used in the ML community today. To address this gap, we present WILDS, a curated benchmark of 10 datasets reflecting a diverse range of distribution shifts that naturally arise in real-world applications, such as shifts across hospitals for tumor identification; across camera traps for wildlife monitoring; and across time and location in satellite imaging and poverty mapping. On each dataset, we show that standard training yields substantially lower out-of-distribution than in-distribution performance. This gap remains even with models trained by existing methods for tackling distribution shifts, underscoring the need for new methods for training models that are more robust to the types of distribution shifts that arise in practice. To facilitate method development, we provide an open-source package that automates dataset loading, contains default model architectures and hyperparameters, and standardizes evaluations. Code and leaderboards are available at https://wilds.stanford.edu.

References in corpus (85)

• Decoupled Weight Decay Regularization
• Neural Message Passing for Quantum Chemistry
• HuggingFace's Transformers: State-of-the-art Natural Language Processing
• Learning Transferable Features with Deep Adaptation Networks
• Deep Domain Confusion: Maximizing for Domain Invariance
• VoxCeleb2: Deep Speaker Recognition
• Shortcut Learning in Deep Neural Networks
• Confounding variables can degrade generalization performance of radiological deep learning models
• A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks
• Frustratingly Easy Domain Adaptation
• Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)
• Machine learning methods for histopathological image analysis
• Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift
• The Open Catalyst 2020 (OC20) Dataset and Community Challenges
• Classifier Technology and the Illusion of Progress
• Mapping the Human Body at Cellular Resolution -- The NIH Common Fund Human BioMolecular Atlas Program
• Solving Rubik's Cube with a Robot Hand
• Quantifying the effects of data augmentation and stain color normalization in convolutional neural networks for computational pathology
• Domain Generalization via Invariant Feature Representation
• The Measure and Mismeasure of Fairness
• Underspecification Presents Challenges for Credibility in Modern Machine Learning
• Domain Generalization via Model-Agnostic Learning of Semantic Features
• Generalisation in humans and deep neural networks
• CodeXGLUE: A Machine Learning Benchmark Dataset for Code Understanding and Generation
• Automatic acoustic detection of birds through deep learning: the first Bird Audio Detection challenge
• Do ImageNet Classifiers Generalize to ImageNet?
• Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness
• Revisiting Batch Normalization For Practical Domain Adaptation
• Learning to Navigate in Complex Environments
• Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization
• Predicting breast tumor proliferation from whole-slide images: the TUPAC16 challenge
• Invariant Risk Minimization
• XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization
• Whole-Slide Mitosis Detection in H&E Breast Histology Using PHH3 as a Reference to Train Distilled Stain-Invariant Convolutional Networks
• Wasserstein Distance Guided Representation Learning for Domain Adaptation
• Tent: Fully Test-time Adaptation by Entropy Minimization
• LEAF: A Benchmark for Federated Settings
• Scalability in Perception for Autonomous Driving: Waymo Open Dataset
• Fairness Without Demographics in Repeated Loss Minimization
• No Classification without Representation: Assessing Geodiversity Issues in Open Data Sets for the Developing World
• Machine learning on DNA-encoded libraries: A new paradigm for hit-finding
• Concept Bottleneck Models
• Predictive Inequity in Object Detection
• Efficient Pipeline for Camera Trap Image Review
• RobustBench: a standardized adversarial robustness benchmark
• Does Distributionally Robust Supervised Learning Give Robust Classifiers?
• Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods
• In Search of Lost Domain Generalization
• An Investigation of Why Overparameterization Exacerbates Spurious Correlations
• Feature Robustness in Non-stationary Health Records: Caveats to Deployable Model Performance in Common Clinical Machine Learning Tasks
• Just Train Twice: Improving Group Robustness without Training Group Information
• SelectiveNet: A Deep Neural Network with an Integrated Reject Option
• Robot Learning in Homes: Improving Generalization and Reducing Dataset Bias
• Noise or Signal: The Role of Image Backgrounds in Object Recognition
• Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment
• Learning from Failure: Training Debiased Classifier from Biased Classifier
• Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English
• Regularized Learning for Domain Adaptation under Label Shifts
• Graph-based, Self-Supervised Program Repair from Diagnostic Feedback
• Dialogue Learning With Human-In-The-Loop
• Neural Program Repair by Jointly Learning to Localize and Repair
• Model Patching: Closing the Subgroup Performance Gap with Data Augmentation
• Semi-supervised Deep Kernel Learning: Regression with Unlabeled Data by Minimizing Predictive Variance
• Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis
• No Subclass Left Behind: Fine-Grained Robustness in Coarse-Grained Classification Problems
• Maximum Likelihood with Bias-Corrected Calibration is Hard-To-Beat at Label Shift Adaptation
• Bias-Reduced Uncertainty Estimation for Deep Neural Classifiers
• TyDi QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages
• Accuracy on the Line: On the Strong Correlation Between Out-of-Distribution and In-Distribution Generalization
• Robustness to Spurious Correlations via Human Annotations
• Evaluating Model Robustness and Stability to Dataset Shift
• Slice-based Learning: A Programming Model for Residual Learning in Critical Data Slices
• An Introduction to Electrocatalyst Design using Machine Learning for Renewable Energy Storage
• Variational Item Response Theory: Fast, Accurate, and Expressive
• The iWildCam 2021 Competition Dataset
• CheXphoto: 10,000+ Photos and Transformations of Chest X-rays for Benchmarking Deep Learning Robustness
• Synthetic Datasets for Neural Program Synthesis
• Evaluating Progress on Machine Learning for Longitudinal Electronic Healthcare Data
• SmartPaste: Learning to Adapt Source Code
• Counting Fish and Dolphins in Sonar Images Using Deep Learning
• Enforcing Predictive Invariance across Structured Biomedical Domains
• Limitations of Pinned AUC for Measuring Unintended Bias
• Fairness and Robustness in Invariant Learning: A Case Study in Toxicity Classification
• Post-Estimation Smoothing: A Simple Baseline for Learning with Side Information
• Empirical Frequentist Coverage of Deep Learning Uncertainty Quantification Procedures

Cited by in corpus (72)

• Domain Generalization: A Survey
• Large Language Models for Code: Security Hardening and Adversarial Testing
• Accountability in an Algorithmic Society: Relationality, Responsibility, and Robustness in Machine Learning
• RobustBench: a standardized adversarial robustness benchmark
• MEMO: Test Time Robustness via Adaptation and Augmentation
• Learning domain-agnostic visual representation for computational pathology using medically-irrelevant style transfer augmentation
• Measuring and modeling the motor system with machine learning
• Sensible AI: Re-imagining Interpretability and Explainability using Sensemaking Theory
• A crowdsourced dataset of aerial images with annotated solar photovoltaic arrays and installation metadata
• Towards a Theoretical Framework of Out-of-Distribution Generalization
• Marrying Fairness and Explainability in Supervised Learning
• Inherent Limitations of AI Fairness
• Exploring the Limits of Large Scale Pre-training
• CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters
• Laplace Redux -- Effortless Bayesian Deep Learning
• Towards out of distribution generalization for problems in mechanics
• Assaying Out-Of-Distribution Generalization in Transfer Learning
• Augmented Datasheets for Speech Datasets and Ethical Decision-Making
• GraphWorld: Fake Graphs Bring Real Insights for GNNs
• SAND-mask: An Enhanced Gradient Masking Strategy for the Discovery of Invariances in Domain Generalization
• MultiBench: Multiscale Benchmarks for Multimodal Representation Learning
• Better, Not Just More: Data-Centric Machine Learning for Earth Observation
• The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning
• Angler: Helping Machine Translation Practitioners Prioritize Model Improvements
• Can Subnetwork Structure be the Key to Out-of-Distribution Generalization?
• Detecting disparities in police deployments using dashcam data
• Laboratory-Scale AI: Open-Weight Models are Competitive with ChatGPT Even in Low-Resource Settings
• Benchmarking Bayesian Deep Learning on Diabetic Retinopathy Detection Tasks
• Designing Ocean Vision AI: An Investigation of Community Needs for Imaging-based Ocean Conservation
• Partial transfusion: on the expressive influence of trainable batch norm parameters for transfer learning
• Gradient Matching for Domain Generalization
• Fairness Evaluation in Text Classification: Machine Learning Practitioner Perspectives of Individual and Group Fairness
• Environment Inference for Invariant Learning
• Size-Invariant Graph Representations for Graph Classification Extrapolations
• An Unbiased Transformer Source Code Learning with Semantic Vulnerability Graph
• Learning Domain Invariant Representations by Joint Wasserstein Distance Minimization
• Did the Model Change? Efficiently Assessing Machine Learning API Shifts
• DORO: Distributional and Outlier Robust Optimization
• SustainBench: Benchmarks for Monitoring the Sustainable Development Goals with Machine Learning
• Shifts: A Dataset of Real Distributional Shift Across Multiple Large-Scale Tasks
• Detecting Errors and Estimating Accuracy on Unlabeled Data with Self-training Ensembles
• Quantifying and Improving Transferability in Domain Generalization
• Auditing Cross-Cultural Consistency of Human-Annotated Labels for Recommendation Systems
• Model-Based Domain Generalization
• Does enforcing fairness mitigate biases caused by subpopulation shift?
• MLDemon: Deployment Monitoring for Machine Learning Systems
• Learning Implicitly with Noisy Data in Linear Arithmetic
• Selective Classification Can Magnify Disparities Across Groups
• A Fine-Grained Analysis on Distribution Shift
• An Information-theoretic Approach to Distribution Shifts
• Adversarial Scrutiny of Evidentiary Statistical Software
• Experimental Standards for Deep Learning in Natural Language Processing Research
• Focus on the Common Good: Group Distributional Robustness Follows
• The iWildCam 2020 Competition Dataset
• Global Wheat Challenge 2020: Analysis of the competition design and winning models
• Types of Out-of-Distribution Texts and How to Detect Them
• Managing ML Pipelines: Feature Stores and the Coming Wave of Embedding Ecosystems
• A Discussion on Building Practical NLP Leaderboards: The Case of Machine Translation
• Robust Generalization of Quadratic Neural Networks via Function Identification
• Sequential Multivariate Change Detection with Calibrated and Memoryless False Detection Rates
• Learning Neural Models for Natural Language Processing in the Face of Distributional Shift
• Two Sides of Meta-Learning Evaluation: In vs. Out of Distribution
• Exploring the Unfairness of DP-SGD Across Settings
• On Language Models for Creoles
• Learning from Discriminatory Training Data
• On the detection of Out-Of-Distribution samples in Multiple Instance Learning
• Image-to-Image Translation of Synthetic Samples for Rare Classes
• Active Assessment of Prediction Services as Accuracy Surface Over Attribute Combinations
• Learning to Extend Program Graphs to Work-in-Progress Code
• On sensitivity of meta-learning to support data
• Are Compressed Language Models Less Subgroup Robust?
• An Empirical Framework for Domain Generalization in Clinical Settings