AI helps you reading Science

AI generates interpretation videos

AI extracts and analyses the key points of the paper to generate videos automatically


pub
Go Generating

AI Traceability

AI parses the academic lineage of this thesis


Master Reading Tree
Generate MRT

AI Insight

AI extracts a summary of this paper


Weibo:
We have demonstrated that domain adaptation problems with unmeasured variables can be recast as covariate shift problems once we obtain samples from c | x, y at train time

Robustness to Spurious Correlations via Human Annotations

ICML, pp.9109-9119, (2020)

Cited by: 5|Views140
EI
Full Text
Bibtex
Weibo

Abstract

The reliability of machine learning systems critically assumes that the associations between features and labels remain similar between training and test distributions. However, unmeasured variables, such as confounders, break this assumption---useful correlations between features and labels at training time can become useless or even h...More

Code:

Data:

0
Introduction
  • The increasing use of machine learning in socioeconomic problems as well as high-stakes decision-making emphasizes the importance of designing models that can perform well over a wide range of users and conditions (Barocas & Selbst, 2016; Blodgett et al, 2016; Hovy & Sgaard, 2015; Tatman, 2017).
  • There exist many more distributional shifts that the designer of a machine learning system may have been unaware of when collecting data, or may be impossible to measure.
  • Existing approaches such as distributional robustness (Ben-Tal et al, 2013; Lam & Zhou, 2015) and domain adaptation (Mansour et al, 2009b; Blitzer et al, 2011; Gong et al, 2013) require either a priori specifying the distribution shifts, or sampling from the target test distributions.
  • Can the authors leverage humans’ strong prior knowledge to understand the possible distribution shifts for a specific machine learning task? The key idea of the paper is to use human commonsense reasoning as a source of information about potential test-time shifts, and effectively use this information to learn robust models
Highlights
  • The increasing use of machine learning in socioeconomic problems as well as high-stakes decision-making emphasizes the importance of designing models that can perform well over a wide range of users and conditions (Barocas & Selbst, 2016; Blodgett et al, 2016; Hovy & Sgaard, 2015; Tatman, 2017)
  • We demonstrate that distributional robustness over unmeasured variables (UV-distributionally robust optimization) results in more robust models that rely less upon spurious correlations
  • We have demonstrated that domain adaptation problems with unmeasured variables can be recast as covariate shift problems once we obtain samples from c | x, y at train time
  • Our UV-distributionally robust optimization approach and experiments show that crowdsourcing can be an effective way of eliciting these unmeasured variables, and we often obtain results close to an oracle model which uses the true c distribution
  • This work is the first step towards explicitly incorporating human knowledge of potential unmeasured variables via natural language annotations, and opens the possibility of methods that make use of counterfactual explanations from domain experts to learn reliable models in high-stakes situations
  • Reproducibility We provide all source code, data, and experiments as part of a worksheet on the CodaLab platform: https://bit.ly/uvdro-codalab
Results
  • The authors demonstrate that distributional robustness over unmeasured variables (UV-DRO) results in more robust models that rely less upon spurious correlations.
  • The authors show that UV-DRO achieves more robust models than baselines as well as other DRO objectives, including that of Duchi et al (2019) (“Covariate Shift DRO”) and Hashimoto et al (2018) (“Baseline DRO”)
Conclusion
  • The authors have demonstrated that domain adaptation problems with unmeasured variables can be recast as covariate shift problems once the authors obtain samples from c | x, y at train time.
  • The authors' UV-DRO approach and experiments show that crowdsourcing can be an effective way of eliciting these unmeasured variables, and the authors often obtain results close to an oracle model which uses the true c distribution.
  • This work is the first step towards explicitly incorporating human knowledge of potential unmeasured variables via natural language annotations, and opens the possibility of methods that make use of counterfactual explanations from domain experts to learn reliable models in high-stakes situations.
Summary
  • Introduction:

    The increasing use of machine learning in socioeconomic problems as well as high-stakes decision-making emphasizes the importance of designing models that can perform well over a wide range of users and conditions (Barocas & Selbst, 2016; Blodgett et al, 2016; Hovy & Sgaard, 2015; Tatman, 2017).
  • There exist many more distributional shifts that the designer of a machine learning system may have been unaware of when collecting data, or may be impossible to measure.
  • Existing approaches such as distributional robustness (Ben-Tal et al, 2013; Lam & Zhou, 2015) and domain adaptation (Mansour et al, 2009b; Blitzer et al, 2011; Gong et al, 2013) require either a priori specifying the distribution shifts, or sampling from the target test distributions.
  • Can the authors leverage humans’ strong prior knowledge to understand the possible distribution shifts for a specific machine learning task? The key idea of the paper is to use human commonsense reasoning as a source of information about potential test-time shifts, and effectively use this information to learn robust models
  • Results:

    The authors demonstrate that distributional robustness over unmeasured variables (UV-DRO) results in more robust models that rely less upon spurious correlations.
  • The authors show that UV-DRO achieves more robust models than baselines as well as other DRO objectives, including that of Duchi et al (2019) (“Covariate Shift DRO”) and Hashimoto et al (2018) (“Baseline DRO”)
  • Conclusion:

    The authors have demonstrated that domain adaptation problems with unmeasured variables can be recast as covariate shift problems once the authors obtain samples from c | x, y at train time.
  • The authors' UV-DRO approach and experiments show that crowdsourcing can be an effective way of eliciting these unmeasured variables, and the authors often obtain results close to an oracle model which uses the true c distribution.
  • This work is the first step towards explicitly incorporating human knowledge of potential unmeasured variables via natural language annotations, and opens the possibility of methods that make use of counterfactual explanations from domain experts to learn reliable models in high-stakes situations.
Related work
  • Related Works and Discussion

    Although our work draws on ideas from domain adaptation (Mansour et al, 2009a;b; Ben-David et al, 2006) causal invariance (Peters et al, 2016; Meinshausen & Buhlmann, 2015; Rothenhausler et al, 2018), and robust optimization (Ben-Tal et al, 2013; Duchi & Namkoong, 2018; Bertsimas et al, 2018; Lam & Zhou, 2015), few prior works seek to elicit information on the possible shifts in unmeasured variables. For example, Heinze-Deml & Meinshausen (2017) improve robustness under unobserved style shifts in images by relying on multiple images which vary only by style. Similarly, Landeiro & Culotta (2016) develop a back-door adjustment which controls for confounding in text classification tasks when the confounder is known. Recent work by Kaushik et al (2019) use crowdsourcing to revise document text given a specific counterfactual label. This approach is orthogonal to our work, as it does not attempt to address shifts in unmeasured variables and is intended to improve models under observed covariate shifts.

    The crowdsourcing aspect of our work builds on existing work on eliciting human commonsense understanding and counterfactual reasoning. Roemmele et al (2011) showed that humans achieve high performance on commonsense causal reasoning tasks, while Sap et al (2019) has used crowdsourcing to build an “if-then” commonsense reasoning dataset. These works support our results which show crowdsourcing can successfully capture how humans reason about unmeasured variables.
Funding
  • MS was additionally supported by the NSF Graduate Research Fellowship Program under Grant No DGE1656518
Reference
  • Angrist, J. D. and Pischke, J. Mostly Harmless Econometrics: An Empiricist’s Companian. Princeton University Press, 2009.
    Google ScholarFindings
  • 104 California Law Review, 3:671–732, 2016.
    Google ScholarFindings
  • Ben-David, S., Blitzer, J., Crammer, K., and Pereira, F. Analysis of representations for domain adaptation. In Advances in Neural Information Processing Systems (NeurIPS), pp. 137–144, 2006.
    Google ScholarLocate open access versionFindings
  • Ben-Tal, A., den Hertog, D., Waegenaere, A. D., Melenberg, B., and Rennen, G. Robust solutions of optimization problems affected by uncertain probabilities. Management Science, 59:341–357, 2013.
    Google ScholarLocate open access versionFindings
  • Bertsimas, D., Gupta, V., and Kallus, N. Data-driven robust optimization. Mathematical Programming Series A, 167, 2018.
    Google ScholarLocate open access versionFindings
  • Blitzer, J., Kakade, S., and Foster, D. P. Domain adaptation with coupled subspaces. In Artificial Intelligence and Statistics (AISTATS), pp. 173–181, 2011.
    Google ScholarLocate open access versionFindings
  • Blodgett, S. L., Green, L., and O’Connor, B. Demographic dialectal variation in social media: A case study of African-American English. In Empirical Methods in Natural Language Processing (EMNLP), pp. 1119–1130, 2016.
    Google ScholarLocate open access versionFindings
  • Buolamwini, J. and Gebru, T. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pp. 77–91, 2018.
    Google ScholarLocate open access versionFindings
  • Duchi, J. and Namkoong, H. Learning models with uniform performance via distributionally robust optimization. arXiv preprint arXiv:1810.08750, 2018.
    Findings
  • Duchi, J., Hashimoto, T., and Namkoong, H. Distributionally robust losses against mixture covariate shifts. https://cs.stanford.edu/̃thashim/assets/publications/condrisk.pdf, 2019.
    Findings
Your rating :
0

 

Tags
Comments
数据免责声明
页面数据均来自互联网公开来源、合作出版商和通过AI技术自动分析结果,我们不对页面数据的有效性、准确性、正确性、可靠性、完整性和及时性做出任何承诺和保证。若有疑问,可以通过电子邮件方式联系我们:report@aminer.cn
小科