Measuring Robustness to Natural Distribution Shifts in Image Classification

Abstract

We study how robust current ImageNet models are to distribution shifts arising from natural variations in datasets. Most research on robustness focuses on synthetic image perturbations (noise, simulated weather artifacts, adversarial examples, etc.), which leaves open how robustness on synthetic distribution shift relates to distribution shift arising in real data. Informed by an evaluation of 204 ImageNet models in 213 different test conditions, we find that there is often little to no transfer of robustness from current synthetic to natural distribution shift. Moreover, most current techniques provide no robustness to the natural distribution shifts in our testbed. The main exception is training on larger and more diverse datasets, which in multiple cases increases robustness, but is still far from closing the performance gaps. Our results indicate that distribution shifts arising in real data are currently an open research problem. We provide our testbed and data as a resource for future work at https://modestyachts.github.io/imagenet-testbed/ .

References in corpus (23)

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• Improved Regularization of Convolutional Neural Networks with Cutout
• Confounding variables can degrade generalization performance of radiological deep learning models
• AutoAugment: Learning Augmentation Policies from Data
• ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness
• Certified Adversarial Robustness via Randomized Smoothing
• Domain Generalization via Invariant Feature Representation
• SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine
• Making Convolutional Networks Shift-Invariant Again
• Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization
• Billion-scale semi-supervised learning for image classification
• Exploring the Landscape of Spatial Robustness
• No Classification without Representation: Assessing Geodiversity Issues in Open Data Sets for the Developing World
• Adversarial Filters of Dataset Biases
• Tencent ML-Images: A Large-Scale Multi-Label Image Database for Visual Representation Learning
• Testing Robustness Against Unforeseen Adversaries
• Wasserstein Adversarial Examples via Projected Sinkhorn Iterations
• YouTube-BoundingBoxes: A Large High-Precision Human-Annotated Data Set for Object Detection in Video
• Using Videos to Evaluate Image Model Robustness
• Robustness to Spurious Correlations via Human Annotations
• Robustness properties of Facebook's ResNeXt WSL models
• Adversarial Training Versus Weight Decay
• Auditing ImageNet: Towards a Model-driven Framework for Annotating Demographic Attributes of Large-Scale Image Datasets

Cited by in corpus (23)

• Learning Transferable Visual Models From Natural Language Supervision
• Learning to Prompt for Vision-Language Models
• Reproducible scaling laws for contrastive language-image learning
• The Fallacy of AI Functionality
• RobustBench: a standardized adversarial robustness benchmark
• Multi-Objective Hyperparameter Optimization in Machine Learning -- An Overview
• ImageNet-Patch: A Dataset for Benchmarking Machine Learning Robustness against Adversarial Patches
• CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters
• Assaying Out-Of-Distribution Generalization in Transfer Learning
• BREEDS: Benchmarks for Subpopulation Shift
• Structure discovery in Atomic Force Microscopy imaging of ice
• An Analysis of Deep Object Detectors For Diver Detection
• (Beyond) Reasonable Doubt: Challenges that Public Defenders Face in Scrutinizing AI in Court
• In-N-Out: Pre-Training and Self-Training using Auxiliary Information for Out-of-Distribution Robustness
• In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models against Variability
• Adaptive transfer learning for surgical tool presence detection in laparoscopic videos through gradual freezing fine-tuning
• Monitoring Robustness and Individual Fairness
• Towards Robust Off-Policy Evaluation via Human Inputs
• Using Synthetic Corruptions to Measure Robustness to Natural Distribution Shifts
• Addressing Weak Decision Boundaries in Image Classification by Leveraging Web Search and Generative Models
• Robust Generalization of Quadratic Neural Networks via Function Identification
• Learning Under Adversarial and Interventional Shifts
• Reappraising Domain Generalization in Neural Networks