Dangers of Bayesian Model Averaging under Covariate Shift

被引:0
|
作者
Izmailov, Pavel [1 ]
Nicholson, Patrick [2 ]
Lotfi, Sanae [1 ]
Wilson, Andrew Gordon [1 ]
机构
[1] NYU, New York, NY 10003 USA
[2] Covera Hlth, New York, NY USA
关键词
ADAPTATION;
D O I
暂无
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Approximate Bayesian inference for neural networks is considered a robust alternative to standard training, often providing good performance on out-of-distribution data. However, Bayesian neural networks (BNNs) with high-fidelity approximate inference via full-batch Hamiltonian Monte Carlo achieve poor generalization under covariate shift, even underperforming classical estimation. We explain this surprising result, showing how a Bayesian model average can in fact be problematic under covariate shift, particularly in cases where linear dependencies in the input features cause a lack of posterior contraction. We additionally show why the same issue does not affect many approximate inference procedures, or classical maximum a-posteriori (MAP) training. Finally, we propose novel priors that improve the robustness of BNNs to many sources of covariate shift.
引用
收藏
页数:14
相关论文
共 50 条
  • [1] On the use of Bayesian model averaging for covariate selection in epidemiological modeling
    Whitney M.
    Ryan L.
    Walkowiak J.
    [J]. Journal of Statistical Theory and Practice, 2013, 7 (2) : 233 - 247
  • [2] Model selection under covariate shift
    Sugiyama, M
    Müller, KR
    [J]. ARTIFICIAL NEURAL NETWORKS: FORMAL MODELS AND THEIR APPLICATIONS - ICANN 2005, PT 2, PROCEEDINGS, 2005, 3697 : 235 - 240
  • [3] Unlabelled Data Improves Bayesian Uncertainty Calibration under Covariate Shift
    Chan, Alex J.
    Alaa, Ahmed M.
    Qian, Zhaozhi
    van der Schaar, Mihaela
    [J]. INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 119, 2020, 119
  • [4] Unlabelled Data Improves Bayesian Uncertainty Calibration under Covariate Shift
    Chan, Alex J.
    Alaa, Ahmed M.
    Qian, Zhaozhi
    van der Schaar, Mihaela
    [J]. 25TH AMERICAS CONFERENCE ON INFORMATION SYSTEMS (AMCIS 2019), 2019,
  • [5] On Prior Selection and Covariate Shift of -Bayesian Prediction Under -Divergence Risk
    Suzuki, Taiji
    Komaki, Fumiyasu
    [J]. COMMUNICATIONS IN STATISTICS-THEORY AND METHODS, 2010, 39 (8-9) : 1655 - 1673
  • [6] Automated covariate selection and Bayesian model averaging in population PK/PD models
    Lunn, David J.
    [J]. JOURNAL OF PHARMACOKINETICS AND PHARMACODYNAMICS, 2008, 35 (01) : 85 - 100
  • [7] Automated covariate selection and Bayesian model averaging in population PK/PD models
    David J. Lunn
    [J]. Journal of Pharmacokinetics and Pharmacodynamics, 2008, 35 : 85 - 100
  • [8] Experimental Bayesian generalization error of non-regular models under covariate shift
    Yamazaki, Keisuke
    Watanabe, Sumio
    [J]. NEURAL INFORMATION PROCESSING, PART I, 2008, 4984 : 466 - 476
  • [9] Training on Test Data with Bayesian Adaptation for Covariate Shift
    Zhou, Aurick
    Levine, Sergey
    [J]. ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 34 (NEURIPS 2021), 2021, 34
  • [10] Robust Fairness under Covariate Shift
    Rezaei, Ashkan
    Liu, Anqi
    Memarrast, Omid
    Ziebart, Brian D.
    [J]. THIRTY-FIFTH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE, THIRTY-THIRD CONFERENCE ON INNOVATIVE APPLICATIONS OF ARTIFICIAL INTELLIGENCE AND THE ELEVENTH SYMPOSIUM ON EDUCATIONAL ADVANCES IN ARTIFICIAL INTELLIGENCE, 2021, 35 : 9419 - 9427