Understanding the Failure Modes of Out-of-Distribution Generalization

Empirical studies suggest that machine learning models often rely on features, such as the background, that may be spuriously correlated with the label only during training time, resulting in poor accuracy during test-time. In this work, we identify the fundamental factors that give rise to this behavior, by explaining why models fail this way {\em even} in easy-to-learn tasks where one would expect these models to succeed. In particular, through a theoretical study of gradient-descent-trained linear classifiers on some easy-to-learn tasks, we uncover two complementary failure modes. These modes arise from how spurious correlations induce two kinds of skews in the data: one geometric in nature, and another, statistical in nature. Finally, we construct natural modifications of image classification datasets to understand when these failure modes can arise in practice. We also design experiments to isolate the two failure modes when training modern neural networks on these datasets.

[1]  Hongyi Zhang,et al.  mixup: Beyond Empirical Risk Minimization , 2017, ICLR.

[2]  Jonas Peters,et al.  Causal inference by using invariant prediction: identification and confidence intervals , 2015, 1501.01332.

[3]  François Laviolette,et al.  Domain-Adversarial Training of Neural Networks , 2015, J. Mach. Learn. Res..

[4]  D. Tao,et al.  Deep Domain Generalization via Conditional Invariant Adversarial Networks , 2018, ECCV.

[5]  David Lopez-Paz,et al.  In Search of Lost Domain Generalization , 2020, ICLR.

[6]  Pietro Perona,et al.  Recognition in Terra Incognita , 2018, ECCV.

[7]  Gilles Blanchard,et al.  Generalizing from Several Related Classification Tasks to a New Unlabeled Sample , 2011, NIPS.

[8]  Ye Xu,et al.  Unbiased Metric Learning: On the Utilization of Multiple Datasets and Web Images for Softening Bias , 2013, 2013 IEEE International Conference on Computer Vision.

[9]  Percy Liang,et al.  An Investigation of Why Overparameterization Exacerbates Spurious Correlations , 2020, ICML.

[10]  Ethan Caballero,et al.  Out-of-Distribution Generalization via Risk Extrapolation (REx) , 2020, ArXiv.

[11]  J. Zico Kolter,et al.  Generalization in Deep Networks: The Role of Distance from Initialization , 2019, ArXiv.

[12]  David Lopez-Paz,et al.  Invariant Risk Minimization , 2019, ArXiv.

[13]  J. Zico Kolter,et al.  Uniform convergence may be unable to explain generalization in deep learning , 2019, NeurIPS.

[14]  Jian Sun,et al.  Deep Residual Learning for Image Recognition , 2015, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[15]  Martin Wattenberg,et al.  Adversarial Spheres , 2018, ICLR.

[16]  Percy Liang,et al.  Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization , 2019, ArXiv.

[17]  Nathan Srebro,et al.  The Implicit Bias of Gradient Descent on Separable Data , 2017, J. Mach. Learn. Res..

[18]  Mahdi Eftekhari,et al.  Towards Shape Biased Unsupervised Representation Learning for Domain Generalization , 2019, ArXiv.

[19]  Matus Telgarsky,et al.  Risk and parameter convergence of logistic regression , 2018, ArXiv.

[20]  Aleksander Madry,et al.  Robustness May Be at Odds with Accuracy , 2018, ICLR.

[21]  Nathan Srebro,et al.  Exploring Generalization in Deep Learning , 2017, NIPS.

[22]  Vladimir Vapnik,et al.  Statistical learning theory , 1998 .

[23]  Bernhard Schölkopf,et al.  Domain Generalization via Invariant Feature Representation , 2013, ICML.

[24]  Jon Howell,et al.  Asirra: a CAPTCHA that exploits interest-aligned manual image categorization , 2007, CCS '07.

[25]  Yongxin Yang,et al.  Learning to Generalize: Meta-Learning for Domain Generalization , 2017, AAAI.

[26]  Hamza Fawzi,et al.  Adversarial vulnerability for any classifier , 2018, NeurIPS.