The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization

Hendrycks, Dan; Basart, Steven; Mu, Norman; Kadavath, Saurav; Wang, Frank; Dorundo, Evan; Desai, Rahul; Zhu, Tyler; Parajuli, Samyak; Guo, Mike; Song, Dawn; Steinhardt, Jacob; Gilmer, Justin

doi:10.1109/iccv48922.2021.00823

Cited by 705 publications

(525 citation statements)

References 13 publications

Supporting

Mentioning

439

Contrasting

Order By: Relevance

“…We extend the work of Wang et al [1] by introducing a novel loss function, using a diversity regularizer, and prepending a parametrized input transformation module to the network. We show that our approach outperform previous works and make pretrained models robust against common corruptions on image classification benchmarks as ImageNet-C [8] and ImageNet-R [9]. Sun et al [10] investigate test-time adaptation using a self-supervision task.…”

Section: Introductionmentioning

confidence: 83%

“…This is beneficial because any good performing pretrained network can be readily reused, e.g., a network trained on some proprietary data not available to the public. We show, that our method significantly improves performance on models that are trained on clean ImageNet data such as a ResNet50 [13], as well as robust models such as ResNet50 models trained using DeepAugment+AugMix [9]. In summary our main contributions are as follows: we propose non-saturating losses based on the negative log likelihood ratio, such that gradients from high confidence predictions still contribute to test-time adaptation.…”

Section: Introductionmentioning

confidence: 92%

“…One such benchmark, ImageNet-C [8], contains simulated corruptions such as noise, blur, weather effects, and digital image transformations. Additionally, Hendrycks et al [9] proposed three data sets containing real-world distribution shifts, including Imagenet-R. The ImageNet-C have been further extended to MNIST [14], several object detection datasets [15], and image segmentation [16], reflecting the interest of the robustness community.…”

Section: Related Workmentioning

confidence: 99%

“…Most proposals for improving robustness involve special training protocols, requiring time and additional resources. This includes data augmentation like Gaussian noise [17,18,9], CutMix [19], AugMix [4], training on stylized images [3,20] or against adversarial noise distributions [21]. Mintun et al [22] pointed out that many improvements on ImageNet-C are due to data augmentations which are too similar to the test corruptions, that is: overfitting to ImageNet-C occurs.…”

Section: Related Workmentioning

confidence: 99%

“…This benchmark also includes 4 additional corruptions as validation data. For domain adaptation, we choose ImageNet trained models to adapt to ImageNet-R proposed by Hendrycks et al [9]. This dataset contains various naturally occurring artistic renditions of object classes from the original ImageNet.…”

Section: Experimental Settingsmentioning

confidence: 99%

See 4 more Smart Citations

Test-Time Adaptation to Distribution Shift by Confidence Maximization and Input Transformation

Mummadi,

Hutmacher,

Rambach

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep neural networks often exhibit poor performance on data that is unlikely under the train-time data distribution, for instance data affected by corruptions. Previous works demonstrate that test-time adaptation to data shift, for instance using entropy minimization [1], effectively improves performance on such shifted distributions. This paper focuses on the fully test-time adaptation setting, where only unlabeled data from the target distribution is required. This allows adapting arbitrary pretrained networks. Specifically, we propose a novel loss that improves test-time adaptation by addressing both premature convergence and instability of entropy minimization. This is achieved by replacing the entropy by a non-saturating surrogate and adding a diversity regularizer based on batch-wise entropy maximization that prevents convergence to trivial collapsed solutions. Moreover, we propose to prepend an input transformation module to the network that can partially undo test-time distribution shifts. Surprisingly, this preprocessing can be learned solely using the fully test-time adaptation loss in an end-to-end fashion without any target domain labels or source domain data. We show that our approach outperforms previous work in improving the robustness of publicly available pretrained image classifiers to common corruptions on such challenging benchmarks as ImageNet-C.Preprint. Under review.

show abstract

Section: Introductionmentioning

confidence: 83%

Section: Introductionmentioning

confidence: 92%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Experimental Settingsmentioning

confidence: 99%

See 3 more Smart Citations

Test-Time Adaptation to Distribution Shift by Confidence Maximization and Input Transformation

Mummadi,

Hutmacher,

Rambach

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

Deep multi‐instance transfer learning for pneumothorax classification in chest X‐ray images

et al. 2021

View full text Add to dashboard Cite

Pneumothorax is a life-threatening emergency that requires immediate treatment. Frontal-view chest X-ray images are typically used for pneumothorax detection in clinical practice. However, manual review of radiographs is time-consuming, labor-intensive, and highly dependent on the experience of radiologists, which may lead to misdiagnosis. Here, we aim to develop a reliable automatic classification method to assist radiologists in rapidly and accurately diagnosing pneumothorax in frontal chest radiographs. Methods: A novel residual neural network (ResNet)-based two-stage deeplearning strategy is proposed for pneumothorax identification: local feature learning (LFL) followed by global multi-instance learning (GMIL). Most of the nonlesion regions in the images are removed for learning discriminative features. Two datasets are used for large-scale validation: a private dataset (27 955 frontal-view chest X-ray images) and a public dataset (the National Institutes of Health [NIH] ChestX-ray14; 112 120 frontal-view X-ray images). The model performance of the identification was evaluated using the accuracy, precision, recall, specificity, F1-score, receiver operating characteristic (ROC), and area under ROC curve (AUC). Fivefold cross-validation is conducted on the datasets, and then the mean and standard deviation of the above-mentioned metrics are calculated to assess the overall performance of the model. Results:The experimental results demonstrate that the proposed learning strategy can achieve state-of -the-art performance on the NIH dataset with an accuracy, AUC, precision, recall, specificity, and F1-score of 94.4% ± 0.7%, 97.3% ± 0.5%, 94.2% ± 0.3%, 94.6% ± 1.5%, 94.2% ± 0.4%, and 94.4% ± 0.7%, respectively. Conclusions: The experimental results demonstrate that our proposed CAD system is an efficient assistive tool in the identification of pneumothorax.

show abstract

Out‐of‐distribution detection with in‐distribution voting using the medical example of chest x‐ray classification

Wollek,

Willem,

Ingrisch

et al. 2023

Medical Physics

View full text Add to dashboard Cite

BackgroundDeep learning models are being applied to more and more use cases with astonishing success stories, but how do they perform in the real world? Models are typically tested on specific cleaned data sets, but when deployed in the real world, the model will encounter unexpected, out‐of‐distribution (OOD) data.PurposeTo investigate the impact of OOD radiographs on existing chest x‐ray classification models and to increase their robustness against OOD data.MethodsThe study employed the commonly used chest x‐ray classification model, CheXnet, trained on the chest x‐ray 14 data set, and tested its robustness against OOD data using three public radiography data sets: IRMA, Bone Age, and MURA, and the ImageNet data set. To detect OOD data for multi‐label classification, we proposed in‐distribution voting (IDV). The OOD detection performance is measured across data sets using the area under the receiver operating characteristic curve (AUC) analysis and compared with Mahalanobis‐based OOD detection, MaxLogit, MaxEnergy, self‐supervised OOD detection (SS OOD), and CutMix.ResultsWithout additional OOD detection, the chest x‐ray classifier failed to discard any OOD images, with an AUC of 0.5. The proposed IDV approach trained on ID (chest x‐ray 14) and OOD data (IRMA and ImageNet) achieved, on average, 0.999 OOD AUC across the three data sets, surpassing all other OOD detection methods. Mahalanobis‐based OOD detection achieved an average OOD detection AUC of 0.982. IDV trained solely with a few thousand ImageNet images had an AUC 0.913, which was considerably higher than MaxLogit (0.726), MaxEnergy (0.724), SS OOD (0.476), and CutMix (0.376).ConclusionsThe performance of all tested OOD detection methods did not translate well to radiography data sets, except Mahalanobis‐based OOD detection and the proposed IDV method. Consequently, training solely on ID data led to incorrect classification of OOD images as ID, resulting in increased false positive rates. IDV substantially improved the model's ID classification performance, even when trained with data that will not occur in the intended use case or test set (ImageNet), without additional inference overhead or performance decrease in the target classification. The corresponding code is available at https://gitlab.lrz.de/IP/a‐knee‐cannot‐have‐lung‐disease.

show abstract

The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization

Cited by 705 publications

References 13 publications

Test-Time Adaptation to Distribution Shift by Confidence Maximization and Input Transformation

Test-Time Adaptation to Distribution Shift by Confidence Maximization and Input Transformation

Deep multi‐instance transfer learning for pneumothorax classification in chest X‐ray images

Out‐of‐distribution detection with in‐distribution voting using the medical example of chest x‐ray classification

Contact Info

Product

Resources

About