Pathologies of Factorised Gaussian and MC Dropout Posteriors in Bayesian Neural Networks

Andrew Y. K. Foong, David R. Burt, Yingzhen Li, Richard E. Turner

Abstract

While Bayesian neural networks (BNNs) hold the promise of being flexible, well-calibrated statistical models, inference often requires approximations whose consequences are poorly understood. We study the quality of common variational methods in approximating the Bayesian predictive distribution. For single-hidden layer ReLU BNNs, we prove a fundamental limitation in function-space of two of the most commonly used distributions defined in weight-space: mean-field Gaussian and Monte Carlo dropout. We find there are simple cases where neither method can have substantially increased uncertainty in between well-separated regions of low uncertainty. We provide strong empirical evidence that exact inference does not have this pathology, hence it is due to the approximation and not the model. In contrast, for deep networks, we prove a universality result showing that there exist approximate posteriors in the above classes which provide flexible uncertainty estimates. However, we find empirically that pathologies of a similar form as in the single-hidden layer case can persist when performing variational inference in deeper networks. Our results motivate careful consideration of the implications of approximate inference methods in BNNs.

Introduction

Bayesian Neural Networks (BNNs) are prized for their theoretical ability to provide well-calibrated uncertainty estimates. However, because exact Bayesian inference is computationally intractable, researchers rely on approximations like Mean-Field Variational Inference (MFVI), which assumes a factorized Gaussian posterior over weights, and Monte Carlo Dropout (MCDO), which uses dropout at test time.

A long-standing observation is that these approximations often fail to show high uncertainty in regions between well-separated data clusters—a property called "in-between uncertainty." This paper investigates whether this failure is a property of the BNN model itself or a limitation of the approximation methods. Understanding this is crucial for safety-critical applications like medical diagnosis or autonomous driving, where knowing what a model doesn't know is vital.

Theoretical Limitations in Shallow Networks

For single-hidden layer (1HL) ReLU BNNs, the authors identify fundamental geometric constraints that prevent MFVI and MCDO from capturing in-between uncertainty.

  1. MFVI Variance Bounds: The variance function for MFVI is bounded by the sum of variances at the endpoints of a line segment. If uncertainty is low at two data points, the approximation mathematically forces the region between them to also have low uncertainty.
  2. MCDO Convexity: For MCDO, the variance function is strictly convex. This implies that the variance cannot have a local maximum—a "bump" of uncertainty—between two clusters of data.
Geometric constraints on variance in shallow MFVI BNNs.

Deep Networks and the Optimization Gap

The authors prove a "universality" result for deep networks (2+ hidden layers), showing that the MFVI and MCDO families are expressive enough to represent arbitrary mean and variance functions. However, this theoretical expressivity does not translate to practical performance when using standard training objectives.

Empirical analysis comparing MFVI and MCDO against "gold-standard" references—Hamiltonian Monte Carlo (HMC) and the infinite-width Gaussian Process (GP) limit—shows that the exact posterior does exhibit high in-between uncertainty. The failure of MFVI and MCDO to match this behavior, even in deep networks, suggests that the standard Evidence Lower Bound (ELBO) objective fails to find the solutions that provide flexible uncertainty estimates.

MFVI and MCDO fail to match the target variance function's central uncertainty 'bump'.

Impact on Active Learning

The practical consequences of these pathologies are demonstrated through an active learning case study on the Naval Propulsion Plant dataset. Because MFVI and MCDO are overconfident in the "middle" of the data distribution, active learning algorithms using them tend to oversample points far from the origin and completely miss important clusters near the center.

On this dataset, active learning with exact inference (GP) reduced RMSE by a factor of three compared to random selection. In contrast, using 1HL MFVI or MCDO actually increased RMSE compared to random selection, showing that poor uncertainty estimates can be worse than no estimates at all.

Conclusion

The study concludes that the overconfidence observed in BNNs is primarily a result of the approximation methods and their optimization objectives rather than the underlying model. While deep networks have the capacity for better uncertainty calibration, current variational methods fail to exploit it. These results motivate the need for more flexible approximating families and improved variational objectives to realize the full potential of Bayesian Neural Networks.

Summary updated

Abstract

While Bayesian neural networks (BNNs) hold the promise of being flexible, well-calibrated statistical models, inference often requires approximations whose consequences are poorly understood. We study the quality of common variational methods in approximating the Bayesian predictive distribution. For single-hidden layer ReLU BNNs, we prove a fundamental limitation in function-space of two of the most commonly used distributions defined in weight-space: mean-field Gaussian and Monte Carlo dropout. We find there are simple cases where neither method can have substantially increased uncertainty in between well-separated regions of low uncertainty. We provide strong empirical evidence that exact inference does not have this pathology, hence it is due to the approximation and not the model. In contrast, for deep networks, we prove a universality result showing that there exist approximate posteriors in the above classes which provide flexible uncertainty estimates. However, we find empirically that pathologies of a similar form as in the single-hidden layer case can persist when performing variational inference in deeper networks. Our results motivate careful consideration of the implications of approximate inference methods in BNNs.
BibTeX
@inproceedings{foong2020pathologies,
  author = {Andrew Y. K. Foong and David R. Burt and Yingzhen Li and Richard E. Turner},
  title = {{Pathologies of Factorised Gaussian and MC Dropout Posteriors in Bayesian Neural Networks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=SI9fhiiNrg5},
}
Figure