A Meta-Analysis of Distributionally-Robust Models

Benjamin Feuer, Ameya Joshi, Chinmay HegdearXiv 2022 · OpenReview

Abstract: State-of-the-art image classifiers trained on massive datasets (such as ImageNet) have been shown to be vulnerable to a range of both intentional and incidental distribution shifts. On the other hand, several recent classifiers with favorable out-of-distribution (OOD) robustness properties have emerged, achieving high accuracy on their target tasks while maintaining their in-distribution accuracy on challenging benchmarks. We present a meta-analysis on a wide range of publicly released models, most of which have been published over the last twelve months. Through this meta-analysis, we empirically identify four main commonalities for all the best-performing OOD-robust models, all of which illuminate the considerable promise of vision-language pre-training.

Introduction: The Brittleness of Modern Classifiers

Deep learning models are notoriously brittle. A classifier that achieves 80% accuracy on the standard ImageNet benchmark might plummet to 30% when shown a sketch of the same object or a photo taken in a different lighting condition. This vulnerability to Out-of-Distribution (OOD) robustness remains a primary hurdle for deploying computer vision systems in the real world.

While many individual papers claim to have improved robustness, the field lacks a unified understanding of which design decisions actually move the needle. This meta-analysis evaluates a representative sample of over 200 publicly available models to answer a fundamental question: What specific properties make an image classifier OOD-robust?

Methodology

To identify the "secret sauce" of robustness, we defined a strict criterion for an "OOD-robust classifier":

  1. High-Accuracy Regime: The model must achieve $\ge 75\%$ on ImageNet.
  2. Relative Robustness: It must maintain $\ge 90\%$ of its in-distribution accuracy when faced with distribution shifts.

Models were categorized based on several covariates: training objective (supervised vs. vision-language contrastive), parameter count, dataset size, and training duration. We evaluated these models across two types of shifts: Natural Distribution Shifts (e.g., ImageNet-A, ImageNet-R, ObjectNet) involving real-world variations, and Artificial Distribution Shifts (e.g., ImageNet-C, Stylized-ImageNet) involving machine-generated corruptions like digital noise.

The Four Pillars of Robustness

Our analysis reveals that OOD robustness is not an accident; it reliably emerges only when four specific design choices co-occur. The single most consistent predictor of natural OOD robustness is the use of a vision-language (VL) contrastive objective (e.g., CLIP). By learning from image-text pairs rather than discrete labels, these models appear to develop a more holistic understanding of visual concepts.

Impact of training covariates on natural OOD robustness.

Beyond the objective, the "Four Pillars" required for top-tier robustness are:

  1. Vision-Language Contrastive Objective: Moving beyond supervised classification.
  2. Massive Datasets: Utilizing $\ge 400M$ samples, often with noisy labels.
  3. High-Capacity Backbones: Architectures like ViT-L or large EfficientNets.
  4. Sufficient Training Duration: Long training schedules measured in many epochs.

State-of-the-Art Performance

Only three models in our entire meta-analysis met the strict definition of being OOD-robust to natural shifts: BASIC, ALIGN, and CLIP-ViT-L-14. These models maintained over 90% of their base accuracy across natural shifts, with BASIC achieving a staggering 97.29% relative natural robustness.

Performance comparison of various models across accuracy regimes and distribution shifts.

However, a "curious counterbalance" exists between natural and artificial robustness. While VL models excel at natural shifts, they often struggle with artificial corruptions like Gaussian noise. Conversely, supervised models trained with "Noisy Student" techniques handle artificial shifts better but fall short on natural ones. Notably, no model achieved the 90% robustness threshold for artificial distribution shifts, suggesting that scaling current VL methods may not be sufficient to overcome digital corruptions.

Conclusion and Future Work

This meta-analysis highlights the immense promise of vision-language pre-training for building resilient vision systems. However, significant challenges remain. Many of the most robust models (like ALIGN and BASIC) rely on massive, private datasets, making it difficult to disentangle the benefits of the algorithm from the diversity of the data. Furthermore, the persistent gap in artificial robustness and the separate challenge of adversarial (intentional) attacks represent critical frontiers for future research. Bridging the gap between natural and artificial OOD performance remains the next major hurdle for the community.

Summary updated

Abstract

State-of-the-art image classifiers trained on massive datasets (such as ImageNet) have been shown to be vulnerable to a range of both intentional and incidental distribution shifts. On the other hand, several recent classifiers with favorable out-of-distribution (OOD) robustness properties have emerged, achieving high accuracy on their target tasks while maintaining their in-distribution accuracy on challenging benchmarks. We present a meta-analysis on a wide range of publicly released models, most of which have been published over the last twelve months. Through this meta-analysis, we empirically identify four main commonalities for all the best-performing OOD-robust models, all of which illuminate the considerable promise of vision-language pre-training.
BibTeX
@misc{feuer2022meta,
  author = {Benjamin Feuer and Ameya Joshi and Chinmay Hegde},
  title = {{A Meta-Analysis of Distributionally-Robust Models}},
  year = {2022},
  eprint = {2206.07565},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.07565},
}
Figure