Dangers of Bayesian Model Averaging under Covariate Shift

Izmailov, Pavel; Nicholson, Patrick; Lotfi, Sanae; Wilson, Andrew Gordon

doi:10.48550/arxiv.2106.11905

Cited by 3 publications

(7 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Finally, Izmailov et al (2021a) show that BNNs have issues under covariate shift, due to the posterior not contracting sufficiently along some directions in the parameter space. The same issue occurs when BNNs are applied to extremely small datasets, which may affect the results on data subsampling presented in Noci et al (2021).…”

Section: Discussionmentioning

confidence: 97%

“…However, there are practical challenges to the adoption of Bayesian deep learning. For example, Izmailov et al (2021a) shows that Bayesian neural networks can profoundly degrade in performance under a wide range of relatively minor distribution shifts -behaviour which could affect applicability on virtually any real-world problem, since train and test rarely come from exactly the same distribution. While their EmpCov prior provides a partial remedy, there is still much work to be done.…”

Section: Discussionmentioning

confidence: 99%

“…Separately from cold posteriors, Izmailov et al (2021a) explain how standard priors for Bayesian neural networks can cause profound deteriorations in performance under covariate shift -with the potential to impact virtually any real-world application of BNNs. They introduce the input dependent EmpCov priors, which helps remedy this issue.…”

Section: Related Workmentioning

confidence: 99%

“…Indeed, q N D (w) does not depend on the labels y, and simply forces the predicted class probabilities to be confident in any one of the classes for the training data. See the EmpCov prior in Izmailov et al (2021a) for another example of a prior that depends on the training inputs but not training labels.…”

Section: Noisy Dirichlet Model: Changing the Prior Over Class Probabi...mentioning

confidence: 99%

See 3 more Smart Citations

On Uncertainty, Tempering, and Data Augmentation in Bayesian Classification

Kapoor¹,

Maddox²,

Izmailov³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Aleatoric uncertainty captures the inherent randomness of the data, such as measurement noise. In Bayesian regression, we often use a Gaussian observation model, where we control the level of aleatoric uncertainty with a noise variance parameter. By contrast, for Bayesian classification we use a categorical distribution with no mechanism to represent our beliefs about aleatoric uncertainty. Our work shows that explicitly accounting for aleatoric uncertainty significantly improves the performance of Bayesian neural networks. We note that many standard benchmarks, such as CIFAR, have essentially no aleatoric uncertainty. Moreover, we show data augmentation in approximate inference has the effect of softening the likelihood, leading to underconfidence and profoundly misrepresenting our honest beliefs about aleatoric uncertainty. Accordingly, we find that a cold posterior, tempered by a power greater than one, often more honestly reflects our beliefs about aleatoric uncertainty than no tempering -providing an explicit link between data augmentation and cold posteriors. We show that we can match or exceed the performance of posterior tempering by using a Dirichlet observation model, where we explicitly control the level of aleatoric uncertainty, without any need for tempering.

show abstract

Section: Discussionmentioning

confidence: 97%

Section: Discussionmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Noisy Dirichlet Model: Changing the Prior Over Class Probabi...mentioning

confidence: 99%

See 2 more Smart Citations

On Uncertainty, Tempering, and Data Augmentation in Bayesian Classification

Kapoor¹,

Maddox²,

Izmailov³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…A popular way to overcome these blemishes is by quantifying (epistemic) uncertainty by aggregating multiple predictions by networks in the Bayesian Model Averaging framework (Jeffreys, 1998;Wilson & Izmailov, 2020), using variational methods (Gal & Ghahramani, 2016;Blundell et al, 2015), ensembling (Lakshminarayanan et al, 2017) or mixtures of the two (Pearce et al, 2020;Wilson & Izmailov, 2020). Nevertheless, many of these methods have been shown not to produce diverse predictions (Wilson & Izmailov, 2020;Fort et al, 2019) and to deliver subpar performance and potentially misleading uncertainty estimates under distributional shift (Ovadia et al, 2019;Masegosa, 2019;Wenzel et al, 2020;Izmailov et al, 2021a;b), raising doubts about their efficacy.…”

Section: Related Workmentioning

confidence: 99%

A Survey on Evidential Deep Learning For Single-Pass Uncertainty Estimation

Ulmer¹

2021

Preprint

View full text Add to dashboard Cite

Popular approaches for quantifying predictive uncertainty in deep neural networks often involve a set of weights or models, for instance via ensembling or Monte Carlo Dropout. These techniques usually produce overhead by having to train multiple model instances or do not produce very diverse predictions. This survey aims to familiarize the reader with an alternative class of models based on the concept of Evidential Deep Learning: For unfamiliar data, they admit "what they don't know" and fall back onto a prior belief. Furthermore, they allow uncertainty estimation in a single model and forward pass by parameterizing distributions over distributions. This survey recapitulates existing works, focusing on the implementation in a classification setting. Finally, we survey the application of the same paradigm to regression problems. We also provide a reflection on the strengths and weaknesses of the mentioned approaches compared to existing ones and provide the most central theoretical results in order to inform future research.

show abstract

Pre-Train Your Loss: Easy Bayesian Transfer Learning with Informative Priors

Shwartz-Ziv¹,

Goldblum²,

Souri³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Deep learning is increasingly moving towards a transfer learning paradigm whereby large foundation models are fine-tuned on downstream tasks, starting from an initialization learned on the source task. But an initialization contains relatively little information about the source task. Instead, we show that we can learn highly informative posteriors from the source task, through supervised or self-supervised approaches, which then serve as the basis for priors that modify the whole loss surface on the downstream task. This simple modular approach enables significant performance gains and more data-efficient learning on a variety of downstream classification and segmentation tasks, serving as a drop-in replacement for standard pre-training strategies. These highly informative priors also can be saved for future use, similar to pre-trained weights, and stand in contrast to the zero-mean isotropic uninformative priors that are typically used in Bayesian deep learning.

show abstract

Dangers of Bayesian Model Averaging under Covariate Shift

Cited by 3 publications

References 0 publications

On Uncertainty, Tempering, and Data Augmentation in Bayesian Classification

On Uncertainty, Tempering, and Data Augmentation in Bayesian Classification

A Survey on Evidential Deep Learning For Single-Pass Uncertainty Estimation

Pre-Train Your Loss: Easy Bayesian Transfer Learning with Informative Priors

Contact Info

Product

Resources

About