Measuring Robustness to Natural Distribution Shifts in Image Classification

By Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, B. Recht, Ludwig Schmidt. 2020

698 citations recorded by the source corpus.

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 this https URL .

Subfields: Artificial Intelligence, Radiology, Nuclear Medicine and Imaging, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Radiology, Nuclear Medicine and Imaging

In Safety, Robustness & Interpretability

Citations in the corpus

Learning Transferable Visual Models From Natural Language Supervision (2021, methodology): “Taori et al. (2020) is a recent comprehensive study moving towards quantifying and understanding these behaviors for ImageNet models.”
On Robustness and Transferability of Convolutional Neural Networks (2020, background): “Concurrently, [61] investigated whether higher accuracy on synthetic datasets translates to superior performance on natural OOD datasets.”
Underspecification Presents Challenges for Credibility in Modern Machine Learning (2020, background): “This brittleness has raised questions about deployments in open-world high-stakes application, and has given rise to an active literature on robustness in image classification (see, e.g., Taori et al., 2020; Djolonga et al., 2020).”
Do Image Classifiers Generalize Across Time? (2019, background): “3 of [28], the authors further analyze the performance of l∞-robust models on”
On Buggy Resizing Libraries and Surprising Subtleties in FID Calculation (2021, background): “Achieving robustness to such perturbations remains an open problem [71], and the preprocessing steps, such as image resizing, used before feature extraction remain consequential.”
CLOOB: Modern Hopfield Networks with InfoLOOB Outperform CLIP (2021, background): “However, modern Hopﬁeld networks lead to a higher similarity of retrieved samples.”
Robust Speech Recognition via Large-Scale Weak Supervision (2022, background): “More recently, Taori et al. (2020) studied the robustness of image classification models, and Miller et al. (2020) investigated this for question-answering models.”
Combined Scaling for Zero-shot Transfer Learning (2021, methodology): “To visualize the data trend, in Figure 3, we plot the accuracy of zero-shot models – BASIC, CLIP (Radford et al., 2021), and ALIGN (Jia et al., 2021) – and of 200 ImageNet-trained models collected by Taori et al. (2020).”
Learning to Prompt for Vision-Language Models (2021, neutral): “Since CoOp requires training on a speciﬁc data distribution, it risks learning spurious correlations that are detrimental to generalization in unseen distributions (domains), as suggested in recent studies (Taori et al., 2020; Zhou et al., 2021).”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, positive): “For example, models can be robust to image corruptions but not to shifts across datasets (Taori et al., 2020; Djolonga et al., 2020), and a method that improves robustness on a standard vision dataset can actually consistently harm robustness on real-world satellite imagery datasets (Xie et al.,…”
Poisoning and Backdooring Contrastive Learning (2021, methodology): “Contrastive models trained on data scraped from the Internet exhibit impressive robustness properties; The contrastively trained CLIP (Radford et al., 2021) model is the ﬁrst technique to show signiﬁcant effective robustness on ImageNet-V2 (Recht et al., 2019; Taori et al., 2020).”
Robust fine-tuning of zero-shot models (2021, methodology): “For the distribution shifts we study, accuracy on the reference distribution is a reliable predictor of accuracy under distribution shift [68,95].”
Are We Learning Yet? A Meta Review of Evaluation Failures Across Machine Learning (2021, methodology): “For example, most models trained on ImageNet were found to experience a considerable drop in accuracy when exposed to images that contained a larger amount of natural variation, such as changes in pose, lighting, object composition, etc. [148].”
On Aliased Resizing and Surprising Subtleties in GAN Evaluation (2022, background): “Achieving robustness to such perturbations remains an open problem [65], and the preprocessing steps, such as image resizing, used before feature extraction remain consequential.”
The Fallacy of AI Functionality (2022, background): “…instance, while many computer vision models developed on ImageNet are tested on synthetic image perturbations in an attempt to measure and improve robustness, but Taori et al. [183] have found that these models are not robust to real-world distribution shifts such as a change in lighting or pose.”
Why Do Better Loss Functions Lead to Less Transferable Features? (2020, result): “By contrast, work that explores out-of-distribution (OOD) generalization, where the training and evaluation datasets consist of the same classes but with some degree of distribution shift, finds that ImageNet accuracy is highly predictive of OOD accuracy across many pretrained models [55, 67].”