Out-of-Distribution Generalization via Risk Extrapolation (REx)

Abstract

Distributional shift is one of the major obstacles when transferring machine learning prediction systems from the lab to the real world. To tackle this problem, we assume that variation across training domains is representative of the variation we might encounter at test time, but also that shifts at test time may be more extreme in magnitude. In particular, we show that reducing differences in risk across training domains can reduce a model's sensitivity to a wide range of extreme distributional shifts, including the challenging setting where the input contains both causal and anti-causal elements. We motivate this approach, Risk Extrapolation (REx), as a form of robust optimization over a perturbation set of extrapolated domains (MM-REx), and propose a penalty on the variance of training risks (V-REx) as a simpler variant. We prove that variants of REx can recover the causal mechanisms of the targets, while also providing some robustness to changes in the input distribution ("covariate shift"). By appropriately trading-off robustness to causally induced distributional shifts and covariate shift, REx is able to outperform alternative methods such as Invariant Risk Minimization in situations where these types of shift co-occur.

References in corpus (32)

• Representation Learning with Contrastive Predictive Coding
• mixup: Beyond Empirical Risk Minimization
• Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor
• Equality of Opportunity in Supervised Learning
• Learning deep representations by mutual information estimation and maximization
• Conditional Adversarial Domain Adaptation
• Understanding deep learning requires rethinking generalization
• AutoAugment: Learning Augmentation Policies from Data
• Contrastive Multiview Coding
• DeepMind Control Suite
• Deep Anomaly Detection with Outlier Exposure
• Do ImageNet Classifiers Generalize to ImageNet?
• Adversarial Examples Are Not Bugs, They Are Features
• Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization
• Certifying Some Distributional Robustness with Principled Adversarial Training
• Invariant Risk Minimization
• Exploring the Landscape of Spatial Robustness
• A Closer Look at Few-shot Classification
• On Learning Invariant Representation for Domain Adaptation
• Detecting and Correcting for Label Shift with Black Box Predictors
• Does Distributionally Robust Supervised Learning Give Robust Classifiers?
• Learning Robust Representations by Projecting Superficial Statistics Out
• Generalizing to unseen domains via distribution matching
• A Fourier Perspective on Model Robustness in Computer Vision
• In Search of Lost Domain Generalization
• Anchor regression: heterogeneous data meets causality
• Natural Neural Networks
• Excessive Invariance Causes Adversarial Vulnerability
• Learning Equations for Extrapolation and Control
• Support and Invertibility in Domain-Invariant Representations
• Fairness risk measures
• Selecting Data Augmentation for Simulating Interventions

Cited by in corpus (70)

• Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learning
• The Risks of Invariant Risk Minimization
• In Search of Lost Domain Generalization
• Deep Long-Tailed Learning: A Survey
• Domain Generalization using Causal Matching
• Data-driven subgrid-scale modeling of forced Burgers turbulence using deep learning with generalization to higher Reynolds numbers via transfer learning
• Understanding the Failure Modes of Out-of-Distribution Generalization
• Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization
• Learning Causal Semantic Representation for Out-of-Distribution Prediction
• Invariance Principle Meets Information Bottleneck for Out-of-Distribution Generalization
• Gradient Starvation: A Learning Proclivity in Neural Networks
• Towards a Theoretical Framework of Out-of-Distribution Generalization
• Self-Supervised Scale Recovery for Monocular Depth and Egomotion Estimation
• Assaying Out-Of-Distribution Generalization in Transfer Learning
• SAND-mask: An Enhanced Gradient Masking Strategy for the Discovery of Invariances in Domain Generalization
• Deep Stable Learning for Out-Of-Distribution Generalization
• Out-of-distribution Prediction with Invariant Risk Minimization: The Limitation and An Effective Fix
• Empirical or Invariant Risk Minimization? A Sample Complexity Perspective
• MT3: Meta Test-Time Training for Self-Supervised Test-Time Adaption
• Learning Causal Models Online
• Invariant Policy Optimization: Towards Stronger Generalization in Reinforcement Learning
• POEM: Polarization of Embeddings for Domain-Invariant Representations
• Better Modelling Out-of-Distribution Regression on Distributed Acoustic Sensor Data Using Anchored Hidden State Mixup
• Learning Invariant Representations and Risks for Semi-supervised Domain Adaptation
• On the Transfer of Disentangled Representations in Realistic Settings
• On Calibration and Out-of-domain Generalization
• Out-of-Distribution Generalization Analysis via Influence Function
• Size-Invariant Graph Representations for Graph Classification Extrapolations
• Latent Causal Invariant Model
• Enforcing Predictive Invariance across Structured Biomedical Domains
• Intervention Design for Effective Sim2Real Transfer
• Adaptive Domain Generalization via Online Disagreement Minimization
• A Theory of Label Propagation for Subpopulation Shift
• Accounting for Unobserved Confounding in Domain Generalization
• Model-Based Domain Generalization
• Understanding Continual Learning Settings with Data Distribution Drift Analysis
• Quantifying and Improving Transferability in Domain Generalization
• Environment Invariant Linear Least Squares
• Vision Based Machine Learning Algorithms for Out-of-Distribution Generalisation
• A Bit More Bayesian: Domain-Invariant Learning with Uncertainty
• Iterative Feature Matching: Toward Provable Domain Generalization with Logarithmic Environments
• Meta-Learned Invariant Risk Minimization
• Linear Regression Games: Convergence Guarantees to Approximate Out-of-Distribution Solutions
• Learning Representations that Support Robust Transfer of Predictors
• Class-conditioned Domain Generalization via Wasserstein Distributional Robust Optimization
• SWAD: Domain Generalization by Seeking Flat Minima
• Multi-label Iterated Learning for Image Classification with Label Ambiguity
• Barycentric-alignment and reconstruction loss minimization for domain generalization
• An Information-theoretic Approach to Distribution Shifts
• Predict then Interpolate: A Simple Algorithm to Learn Stable Classifiers
• Learning Stable Classifiers by Transferring Unstable Features
• Learning Diverse Representations for Fast Adaptation to Distribution Shift
• Optimization-based Causal Estimation from Heterogenous Environments
• Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning
• Learning Under Adversarial and Interventional Shifts
• TAMPC: A Controller for Escaping Traps in Novel Environments
• Explainability-aided Domain Generalization for Image Classification
• Causal Learning for Socially Responsible AI
• Neural Networks for Learning Counterfactual G-Invariances from Single Environments
• Out-of-Distribution Generalization in Kernel Regression
• On Invariance Penalties for Risk Minimization
• Encouraging Intra-Class Diversity Through a Reverse Contrastive Loss for Better Single-Source Domain Generalization
• Robust Domain-Free Domain Generalization with Class-aware Alignment
• An Empirical Framework for Domain Generalization in Clinical Settings
• Domain Conditional Predictors for Domain Adaptation
• Achieving Adversarial Robustness Requires An Active Teacher
• Finding Valid Adjustments under Non-ignorability with Minimal DAG Knowledge
• Learning Less Generalizable Patterns with an Asymmetrically Trained Double Classifier for Better Test-Time Adaptation
• Pareto Adversarial Robustness: Balancing Spatial Robustness and Sensitivity-based Robustness
• Counterfactual Supervision-based Information Bottleneck for Out-of-Distribution Generalization