Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization

Sagawa, Shiori; Koh, Pang Wei; Hashimoto, Tatsunori; Liang, Percy

doi:10.48550/arxiv.1911.08731

Cited by 171 publications

(314 citation statements)

References 39 publications

Supporting

Mentioning

309

Contrasting

Order By: Relevance

“…Empirically, we conduct studies on a set of challenging synthetic linear benchmarks designed by (Aubin et al, 2021) and a set of real-world datasets (two image datasets and one text dataset) used in Sagawa et al (2019). Our empirical results on the synthetic benchmarks validate the claimed environment complexities, and also demonstrate its superior performance when compared with IRM and its variant.…”

Section: Introductionmentioning

confidence: 68%

“…Since the real-world data are highly complex and non-linear, over which the ISR approach cannot be directly applied, we apply ISR on top of the features extracted by the hidden layers of trained neural nets as a post-processing procedure. Experiments show that ISR-Mean can consistently increase the worse-case accuracy of the trained models against spurious correlations and group shifts, and this includes models trained by ERM, reweighting and GroupDRO (Sagawa et al, 2019).…”

Section: Introductionmentioning

confidence: 91%

“…The simplest approach for DG is empirical risk minimization (Vapnik, 1992), which minimizes the sum of empirical risks over all training environments. Distributionally robust optimization is another approach (Sagawa et al, 2019;Volpi et al, 2018), which optimizes models over a worst-case distribution that is perturbed around the original distribution. Besides, there are two popular approaches, domain-invariant representation learning and invariant risk minimization, which we will discuss in detail below.…”

Section: Related Workmentioning

confidence: 99%

“…We adopt three datasets that Sagawa et al (2019) proposes to study the robustness of models against spurious correlations and group shifts. See Fig.…”

Section: Real Datasetsmentioning

confidence: 99%

“…Waterbirds (Sagawa et al, 2019): This is a image dataset built from the CUB (Wah et al, 2011) and Places (Zhou et al, 2017) datasets. The task of this dataset is the classification of waterbirds vs. landbirds.…”

Section: Real Datasetsmentioning

confidence: 99%

See 4 more Smart Citations

Provable Domain Generalization via Invariant-Feature Subspace Recovery

Wang¹,

Si²,

Li³

et al. 2022

Preprint

View full text Add to dashboard Cite

Domain generalization asks for models trained on a set of training environments to perform well on unseen test environments. Recently, a series of algorithms such as Invariant Risk Minimization (IRM) has been proposed for domain generalization. However, Rosenfeld et al. (2021) shows that in a simple linear data model, even if non-convexity issues are ignored, IRM and its extensions cannot generalize to unseen environments with less than d s `1 training environments, where d s is the dimension of the spuriousfeature subspace. In this paper, we propose to achieve domain generalization with Invariantfeature Subspace Recovery (ISR). Our first algorithm, ISR-Mean, can identify the subspace spanned by invariant features from the first-order moments of the class-conditional distributions, and achieve provable domain generalization with d s `1 training environments under the data model of Rosenfeld et al. (2021). Our second algorithm, ISR-Cov, further reduces the required number of training environments to Op1q using the information of second-order moments. Notably, unlike IRM, our algorithms bypass non-convexity issues and enjoy global convergence guarantees. Empirically, our ISRs can obtain superior performance compared with IRM on synthetic benchmarks. In addition, on three real-world image and text datasets, we show that ISR-Mean can be used as a simple yet effective post-processing method to increase the worst-case accuracy of trained models against spurious correlations and group shifts. The code is released at https: //github.com/Haoxiang-Wang/ISR.

show abstract

Section: Introductionmentioning

confidence: 68%

Section: Introductionmentioning

confidence: 91%

Section: Related Workmentioning

confidence: 99%

“…We adopt three datasets that Sagawa et al (2019) proposes to study the robustness of models against spurious correlations and group shifts. See Fig.…”

Section: Real Datasetsmentioning

confidence: 99%

Section: Real Datasetsmentioning

confidence: 99%

See 3 more Smart Citations

Provable Domain Generalization via Invariant-Feature Subspace Recovery

Wang¹,

Si²,

Li³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

Microsoft Azure IoT Platform

Bansal

2020

Designing Internet of Things Solutions With Microsoft Azure

View full text Add to dashboard Cite

To address the problem of NLP classifiers learning spurious correlations between training features and target labels, a common approach is to make the model's predictions invariant to these features. However, this can be counterproductive when the features have a non-zero causal effect on the target label and thus are important for prediction. Therefore, using methods from the causal inference literature, we propose an algorithm to regularize the learnt effect of the features on the model's prediction to the estimated effect of feature on label. This results in an automated augmentation method that leverages the estimated effect of a feature to appropriately change the labels for new augmented inputs. On toxicity and IMDB review datasets, the proposed algorithm minimises spurious correlations and improves the minority group (i.e., samples breaking spurious correlations) accuracy, while also improving the total accuracy compared to standard training. 1

show abstract

RealPatch: A Statistical Matching Framework for Model Patching with Real Samples

Romiti

Inskip

Sharmanska

et al. 2022

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Machine learning classifiers are typically trained to minimise the average error across a dataset. Unfortunately, in practice, this process often exploits spurious correlations caused by subgroup imbalance within the training data, resulting in high average performance but highly variable performance across subgroups. Recent work to address this problem proposes model patching with CAMEL. This previous approach uses generative adversarial networks to perform intra-class inter-subgroup data augmentations, requiring (a) the training of a number of computationally expensive models and (b) sufficient quality of model's synthetic outputs for the given domain. In this work, we propose RealPatch, a framework for simpler, faster, and more data-efficient data augmentation based on statistical matching. Our framework performs model patching by augmenting a dataset with real samples, mitigating the need to train generative models for the target task. We demonstrate the effectiveness of Re-alPatch on three benchmark datasets, CelebA, Waterbirds and a subset of iWildCam, showing improvements in worst-case subgroup performance and in subgroup performance gap in binary classification. Furthermore, we conduct experiments with the imSitu dataset with 211 classes, a setting where generative model-based patching such as CAMEL is impractical. We show that RealPatch can successfully eliminate dataset leakage while reducing model leakage and maintaining high utility. The code for RealPatch can be found at https://github.com/wearepal/RealPatch.

show abstract

Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization

Cited by 171 publications

References 39 publications

Provable Domain Generalization via Invariant-Feature Subspace Recovery

Provable Domain Generalization via Invariant-Feature Subspace Recovery

Microsoft Azure IoT Platform

RealPatch: A Statistical Matching Framework for Model Patching with Real Samples

Contact Info

Product

Resources

About