Tent: Fully Test-time Adaptation by Entropy Minimization

Abstract

A model must adapt itself to generalize to new and different data during testing. In this setting of fully test-time adaptation the model has only the test data and its own parameters. We propose to adapt by test entropy minimization (tent): we optimize the model for confidence as measured by the entropy of its predictions. Our method estimates normalization statistics and optimizes channel-wise affine transformations to update online on each batch. Tent reduces generalization error for image classification on corrupted ImageNet and CIFAR-10/100 and reaches a new state-of-the-art error on ImageNet-C. Tent handles source-free domain adaptation on digit recognition from SVHN to MNIST/MNIST-M/USPS, on semantic segmentation from GTA to Cityscapes, and on the VisDA-C benchmark. These results are achieved in one epoch of test-time optimization without altering training.

arXiv commentICLR 2021 Spotlight

References in corpus (15)

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition
• How transferable are features in deep neural networks?
• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
• The Cityscapes Dataset for Semantic Urban Scene Understanding
• ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness
• Model Adaptation: Unsupervised Domain Adaptation without Source Data
• Adversarial Discriminative Domain Adaptation
• A DIRT-T Approach to Unsupervised Domain Adaptation
• Generalisation in humans and deep neural networks
• Do ImageNet Classifiers Generalize to ImageNet?
• Revisiting Batch Normalization For Practical Domain Adaptation
• Deep High-Resolution Representation Learning for Visual Recognition
• Evaluating Prediction-Time Batch Normalization for Robustness under Covariate Shift

Cited by in corpus (44)

• Domain Generalization: A Survey
• WILDS: A Benchmark of in-the-Wild Distribution Shifts
• A Comprehensive Survey on Test-Time Adaptation under Distribution Shifts
• Synthetic Datasets for Autonomous Driving: A Survey
• Unsolved Problems in ML Safety
• MEMO: Test Time Robustness via Adaptation and Augmentation
• Uncertainty-Induced Transferability Representation for Source-Free Unsupervised Domain Adaptation
• Test-Time Adaptation to Distribution Shift by Confidence Maximization and Input Transformation
• Enhancing and Adapting in the Clinic: Source-free Unsupervised Domain Adaptation for Medical Image Enhancement
• Adversarial purification with Score-based generative models
• Test-time Batch Statistics Calibration for Covariate Shift
• SALUDA: Surface-based Automotive Lidar Unsupervised Domain Adaptation
• MixNorm: Test-Time Adaptation Through Online Normalization Estimation
• Minimax Active Learning
• Anatomy-guided domain adaptation for 3D in-bed human pose estimation
• Test time Adaptation through Perturbation Robustness
• Test-Time Personalization with a Transformer for Human Pose Estimation
• On Interaction Between Augmentations and Corruptions in Natural Corruption Robustness
• Adaptive Denoising via GainTuning
• PIG: Prompt Images Guidance for Night-Time Scene Parsing
• Taking it further: leveraging pseudo labels for field delineation across label-scarce smallholder regions
• Mixture Domain Adaptation to Improve Semantic Segmentation in Real-World Surveillance
• Adaptive Domain Generalization via Online Disagreement Minimization
• Fighting Gradients with Gradients: Dynamic Defenses against Adversarial Attacks
• Online Distillation with Continual Learning for Cyclic Domain Shifts
• Towards Foundation Models and Few-Shot Parameter-Efficient Fine-Tuning for Volumetric Organ Segmentation
• A Bayesian Approach to OOD Robustness in Image Classification
• Distribution Aligned Semantics Adaption for Lifelong Person Re-Identification
• Certified Adversarial Defenses Meet Out-of-Distribution Corruptions: Benchmarking Robustness and Simple Baselines
• Semi-Supervised Hypothesis Transfer for Source-Free Domain Adaptation
• Re-using Adversarial Mask Discriminators for Test-time Training under Distribution Shifts
• Cross-Domain Sentiment Classification with In-Domain Contrastive Learning
• Training on Test Data with Bayesian Adaptation for Covariate Shift
• Generalizable Model-agnostic Semantic Segmentation via Target-specific Normalization
• Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration
• Encoding Robustness to Image Style via Adversarial Feature Perturbations
• Compositional Models: Multi-Task Learning and Knowledge Transfer with Modular Networks
• RoMA: Robust Model Adaptation for Offline Model-based Optimization
• Towards Adversarial Robustness via Transductive Learning
• Distribution Mismatch Correction for Improved Robustness in Deep Neural Networks
• Physically Interpretable Probabilistic Domain Characterization
• Efficient Test Time Adapter Ensembling for Low-resource Language Varieties
• On-target Adaptation
• Towards Data-Free Domain Generalization