Bayesian inference is especially compelling for deep neural networks. The key
distinguishing property of a Bayesian approach is marginalization instead of
optimization, not the prior, or Bayes rule. Neural networks are typically
underspecified by the data, and can represent many different but high performing
models corresponding to different settings of parameters, which is exactly when
marginalization will make the biggest difference for accuracy and calibration.
Moreover, deep ensembles can be seen as approximate Bayesian marginalization.
In many situations, the predictive distribution we want to compute is given by
\begin{equation*}
p(y|x, \mathcal{D}) = \int p(y|x, w) p(w|\mathcal{D}) dw \,.
\end{equation*}
The outputs are $$y$$ (e.g., class labels, regression values, … ), indexed by inputs $$x$$ (e.g. images,
spatial locations, …), the weights (or parameters) of the model $$f(x; w)$$ are $$w$$, and $$D$$ are the
data. Eq. (1) represents a Bayesian model average (BMA). Rather than bet everything on one
hypothesis with a single setting of parameters $$w$$ we want to use every possible setting of
parameters, weighted by their posterior probabilities. This process is called marginalization
of the parameters $$w$$, since the predictive distribution of interest no longer conditions on
$$w$$. This is not a controversial equation, but a direct expression of the sum and product
rules of probability. The BMA represents epistemic uncertainty that is, uncertainty over
which setting of weights (hypothesis) is correct, given limited data. Epistemic uncertainty
is sometimes referred to as model uncertainty, in contrast to aleatoric uncertainty coming
from noise in the measurement process. One can naturally visualize epistemic uncertainty in
regression, by looking at the spread of the predictive distribution as we move in $$x$$ space. As
we move away from the data, there are many more functions that are consistent with our
observations, and so our epistemic uncertainty should grow.
In classical training, one typically finds the regularized maximum likelihood solution
\begin{equation*}
\hat w = \arg\max_w \log p(w|\mathcal{D}) = \arg\max_w (\log p(\mathcal{D}|w) + \log p(w) + \text{constant}).
\end{equation*}
This procedure is sometimes called maximum a-posteriori (MAP) optimization, as it involves maximizing a posterior.
$$\log p(\mathcal{D}|w)$$ is the log likelihood, formed by relating the
function we want to learn $$f(x;w)$$ to our observations.
If we are performing classification with a softmax link function, $$-\log p(\mathcal{D}|w)$$ corresponds to the
cross entropy loss. If we are are performing regression with Gaussian noise, such that
$$p(\mathcal{D}|w) = \prod_{j=1}^{n} p(y_j | w, x_j) = \prod_{j=1}^{n} \mathcal{N}(y_j; f(x_i;w),\sigma^2)$$, then $$-\log p(\mathcal{D}|w)$$
is a scaled MSE loss. In this context,
the prior p(w) acts as a regularizer.
If we choose a flat prior, which has no preference for any setting of the parameters $$w$$ (it does not assign any feasible
setting any more prior density than any other), then it will have no effect on the optimization solution.
On the other hand, a flat prior may have a major effect on marginalization. Indeed, even though
MAP involves a posterior and a prior, and an instantiation of Bayes rule, it is
not at all Bayesian, since it is performing optimization to bet everything on a single hypothesis
$$f(x;\hat{w})$$.
We can view classical training as performing approximate Bayesian inference, using the approximate posterior
$$p(w | \mathcal{D}) \approx \delta(w=\hat{w})$$, where $$\delta$$ is a Dirac delta function that is zero everywhere
except at $$\hat{w}$$. In this case, we recover the standard predictive distribution $$p(y|x,\hat{w})$$. From this perspective,
many alternatives, albeit imperfect, will be preferable
hypothesis with a single setting of parameters $$w$$ we want to use every possible setting of
including impoverished Gaussian posterior approximations for $$p(w|\mathcal{D})$$,
even if the posterior or likelihood are actually highly non-Gaussian and multimodal.
The difference between a classical and Bayesian approach will depend on how sharply peaked the posterior $$p(w|\mathcal{D})$$
becomes. If the posterior is sharply peaked, there may be almost no difference, since a point mass may then be a reasonable approximation
of the posterior.
However, deep neural networks are typically very underspecified by the available data, and will thus have diffuse likelihoods $$p(\mathcal{D}|w)$$.
Not only are
the likelihoods diffuse, but different settings of the parameters correspond to a diverse variety of compelling explanations for the data.
Indeed, Garipov et al. [5] shows that there are large valleys in the loss landscape of neural networks, over which parameters incur
very little loss, but give rise to high performing functions which make meaningfully different predictions on test data. Zona et al. [33]
also demonstrates the variety of good solutions that can be expressed by a neural network posterior.
This is exactly the setting when we most want to perform a Bayesian model average, which will lead to an ensemble containing
many different but high performing models, for better accuracy and better calibration than classical training.
The recent success of deep ensembles [12] is not discouraging, but indeed strong motivation for
following a Bayesian approach.
Deep ensembles involves MAP training of the same architecture many times starting from different random initializations, to find
different local optima.
Thus using these models in an ensemble is an approximate Bayesian model average, with weights that correspond to
models with high likelihood and diverse predictions.
Instead of using a single point mass to approximate our posterior, as with classical training, we are now using multiple
point masses in good locations, enabling a better approximation to the integral in Eq.~(1) that we are trying to solve. The functional
diversity is important for a good approximation to the BMA integral, because we are summing together terms of the form $$p(y|x,w)$$; if two settings of
the weights $$w_i$$ and $$w_j$$ each provide high likelihood (and consequently high posterior mass), but give rise to similar models, then they will be
largely redundant in the model average, and the second setting of parameters will not contribute much to the integral estimate.
While a recent
report [22] shows that deep ensembles appear to outperform some particular approaches to Bayesian neural networks, there are
two key reasons behind these results that are actually optimistic for Bayesian approaches. First, the deep ensembles being used
are finding many different basins of attraction, corresponding to diverse solutions, which enables a better approximation to a Bayesian model
average than the specific Bayesian methods considered in Ovadia et al. [22], which focus their modelling effort on a single
basin of attraction. The second is that the deep ensembles require retraining a network from scratch many times, which incurs a great computational
expense. If one were to control for computation, the approaches which focus on a single basin may be preferred.
There is an important distinction between a Bayesian model average and some approaches to ensembling. The Bayesian model average
assumes that one hypothesis (one setting of the weights) is correct, and averages over models due to an inability to distinguish
between hypotheses given limited data [19]. As we observe more data, the posterior collapses, and the Bayesian model average converges
to the maximum likelihood solution. If the true explanation for the data is actually a combination of hypotheses, the Bayesian model
average may then perform worse as we observe more data. Some ensembling methods instead work by enriching the hypothesis space, and thus
do not collapse in this way. Deep ensembles, however, are finding different MAP or maximum likelihood solutions, corresponding to different
basins of attraction, starting from different random initializations. Therefore the deep ensemble will collapse when the posterior concentrates,
as with a Bayesian model average. Since the hypothesis space is highly expressive for a modern neural network, posterior collapse in many
cases is desirable.
Regarding priors, the prior that matters is the prior in function space, not parameter space. In the case of a Gaussian process
[e.g. 28],
a vague prior would be disastrous, as it is a prior directly in function space and would correspond to white noise. However, when we
combine a vague prior over parameters $$p(w)$$ with a structured function form $$f(x;w)$$ such as a convolutional neural network (CNN),
we induce a structured prior distribution over functions $$p(f(x;w))$$. Indeed, the inductive biases and equivariance
constraints in such models is why they work well in classical settings. We can sample from this induced prior over functions by first sampling parameters
from $$p(w)$$ and then conditioning on these parameters in $$f(x;w)$$ to form a sample from $$p(f(x;w))$$ [e.g., 29, Ch 2].
Alternatively, we can use a neural network kernel with a Gaussian process, to induce a structured distribution over functions [30].
Bayesian or not, the prior, just like the functional form of a model, or the likelihood, will certainly be imperfect,
and making unassailable assumptions will be impossible. Attempting to avoid an important part of the modelling process
because one has to make assumptions, however, will often be a worse alternative than an imperfect assumption.
There are many considerations one might have in
selecting a prior. Sometimes a consideration is invariance under reparametrization.
Parametrization invariance is also a question in considering
regularizers, optimization procedures, model specification, etc., and is not specific to
whether or not one should follow a Bayesian approach. Nonetheless, I will make some
brief additional remarks on these questions.
If we truly have a vague prior over parameters, perhaps subject to some constraint for normalization, then our posterior reflects essentially
the same relative preferences between models as our likelihood, for it is a likelihood scaled by a factor that does not depend on $$w$$ outside some
basic constraints.
In computing the integral for a Bayesian model average, each setting of parameters is weighted by the quality of the associated function, as measured
by the likelihood. Thus the model average is happening in function space, and is invariant to reparametrization. In the context of many standard
architectural specifications,
there are also some additional
benefits to using relatively broad zero-mean centred Gaussian priors, which help provide smoothness in function space by bounding the norm of the weights. But this smoothness is not
a central reason to follow a Bayesian approach, as one could realize similar advantages in performing MAP
optimization. Bayesian methods are fundamentally about marginalization as an alternative to optimization.
Moreover, vague priors over parameters are also often a reasonable description of our a priori subjective beliefs. We want to use a given functional form, which is
by no means vague, but we often do not
have any strong a priori preference for a setting of the parameters.
It is worth
reiterating that a vague prior in parameter space combined with a highly structured model such as a convolutional neural network does not
imply a vague prior in function space, which is also why classical training of neural networks provides good results.
Indeed, vague parameter priors are often preferable to entirely ignoring
epistemic uncertainty, which would be the standard alternative.
In fact, ignoring epistemic uncertainty is a key reason that standard neural network training is miscalibrated. By erroneously
assuming that the model (parameter setting we want to use) is completely determined by a finite dataset, the predictive distribution becomes overconfident: for example, the highest softmax
output of a CNN that has undergone standard training (e.g. MAP optimization) will typically be much higher than the probability of the corresponding class label
[7].
Importantly, ignoring epistemic uncertainty also leads to worse accuracy in point predictions, because we are now ignoring all the other compelling explanations
for the data. While improvements in calibration are an empirically recognized benefit of a Bayesian approach, the enormous potential for
gains in accuracy through Bayesian marginalization with neural networks is a largely overlooked advantage.
There are also many examples where flat priors over parameters
combined with marginalization sidestep pathologies of maximum likelihood training.
Priors without marginalization are
simply regularization, but Bayesian methods are not about regularization [17, Ch 28].
And there is a large body of work considering approximate Bayesian
methods with uninformative priors over parameters (but not functions)
[e.g., 3, 2, 21, 1, 6, 17, 14, 20].
This approach is well-motivated, marginalization is still compelling, and the results are often better than regularized optimization.
By accounting for epistemic uncertainty through uninformative parameter (but not function) priors, we, as a community, have developed
Bayesian deep learning methods with improved calibration, reliable predictive distributions, and improved accuracy
[e.g., 15, 20, 4, 25, 10, 24, 11, 18, 27, 9, 32]. MacKay [16] and Neal [20]
are particularly notable early works considering Bayesian inference for neural networks.
Seeger [26] also provides a clear tutorial on Bayesian methods in machine learning.
Of course, we can always make better assumptions Bayesian or not. We should strive to build more interpretable parameter priors. There
are works that consider building more informative parameter priors for neural networks by reasoning in function space
[e.g., 27, 31, 13, 8].
And we should also build better posterior approximations. Deep ensembles are a promising step in this direction.
But we should not undermine the progress we are making so far. Bayesian inference is especially compelling for deep neural networks. Bayesian deep learning
is gaining visibility because we are making progress, with good and increasingly scalable practical results. We should not discourage these efforts.
If we are shying away from an approximate Bayesian approach because of some challenge or imperfection, we should always ask, “what is the alternative”? The alternative may
indeed be a more impoverished representation of the predictive distribution we want to compute.
There are certainly many challenges to computing the integral of Eq. (1) for modern neural networks, including a posterior landscape which is
difficult to navigate, and an enormously high (e.g., 30 million) dimensional parameter space. Many of the above papers are working towards addressing such
challenges.
We have been particularly working on
recycling geometric information in the SGD trajectory for scalable approximate Bayesian inference
[9, 18], exploiting large loss valleys [5], and creating subspaces of low dimensionality
that capture much of the variability of the network [9]. Pradier et al. [23] also considers different approaches
to dimensionality reduction, based on non-linear transformations.
For exploring multiple distinct basins of attraction, we have been developing
cyclical stochastic MCMC approaches [32], which could be seen as sharing many of the advantages of deep ensembles, but with an added attempt to also marginalize
within basins of attraction.
If you wish to cite this note in your scientific work, here is the bibliographic information:@article{wilson2019bayesian, title={The Case for {B}ayesian Deep Learning}, author={Andrew Gordon Wilson}, note={Accessible at https://cims.nyu.edu/~andrewgw/caseforbdl.pdf}, journal={NYU Courant Technical Report}, year={2019} } A PDF version is available here.[1] James Berger et al. The case for objective Bayesian analysis. Bayesian analysis, 1(3):
385402, 2006.
[2] James O Berger and Luis R Pericchi. The intrinsic Bayes factor for model selection and
prediction. Journal of the American Statistical Association, 91(433):109122, 1996.
[3] Merlise Clyde and Edward I George. Model uncertainty. Statistical science, pages 8194,
2004.
[4] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing
model uncertainty in deep learning. In international conference on machine learning,
pages 10501059, 2016.
[5] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and
Andrew Gordon Wilson. Loss surfaces, mode connectivity, and fast ensembling of DNNs. In Neural
Information Processing Systems, 2018.
[6] Andrew Gelman, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and
Donald B Rubin. Bayesian data analysis. Chapman and Hall/CRC, 2013.
[7] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern
neural networks. In Proceedings of the 34th International Conference on Machine
Learning-Volume 70, pages 13211330. JMLR. org, 2017.
[8] Danijar Hafner, Dustin Tran, Alex Irpan, Timothy Lillicrap, and James Davidson.
Reliable uncertainty estimates in deep neural networks using noise contrastive priors.
arXiv preprint arXiv:1807.09289, 2018.
[9] Pavel Izmailov, Wesley J Maddox, Polina Kirichenko, Timur Garipov, Dmitry Vetrov,
and Andrew Gordon Wilson. Subspace inference for Bayesian deep learning. In
Uncertainty in Artificial Intelligence, 2019.
[10] Alex Kendall and Yarin Gal. What uncertainties do we need in Bayesian deep learning
for computer vision? In Advances in neural information processing systems, pages
55745584, 2017.
[11] Mohammad Emtiyaz Khan, Didrik Nielsen, Voot Tangkaratt, Wu Lin, Yarin Gal, and
Akash Srivastava. Fast and scalable bayesian deep learning by weight-perturbation in
adam. arXiv preprint arXiv:1806.04854, 2018.
[12] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and
scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural
Information Processing Systems, pages 64026413, 2017.
[13] Christos Louizos, Xiahan Shi, Klamer Schutte, and Max Welling. The functional neural
process. In Advances in Neural Information Processing Systems, 2019.
[14] D. J. MacKay. Bayesian interpolation. Neural Computation, 4(3):415447, 1992.
[15] David JC MacKay. Bayesian methods for adaptive models. PhD thesis, California
Institute of Technology, 1992.
[16] David JC MacKay. Probable networks and plausible predictions?a review of practical
bayesian methods for supervised neural networks. Network: computation in neural
systems, 6(3):469505, 1995.
[17] David JC MacKay. Information theory, inference and learning algorithms. Cambridge
university press, 2003.
[18] Wesley Maddox, Timur Garipov, Pavel Izmailov, Dmitry Vetrov, and Andrew Gordon
Wilson. A simple baseline for bayesian uncertainty in deep learning. In Advances in
Neural Information Processing Systems, 2019.
[19] Thomas P Minka. Bayesian model averaging is not model combination. Available
electronically athttp://www.stat.cmu.edu/minka/papers/bma.html, 2000.
[20] R.M. Neal. Bayesian Learning for Neural Networks. Springer Verlag, 1996. ISBN
0387947248.
[21] Anthony OHagan. Fractional Bayes factors for model comparison. Journal of the Royal
Statistical Society: Series B (Methodological), 57(1):99118, 1995.
[22] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, D Sculley, Sebastian Nowozin,
Joshua V Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your
models uncertainty? evaluating predictive uncertainty under dataset shift.
arXiv preprint arXiv:1906.02530, 2019.
[23] Melanie F Pradier, Weiwei Pan, Jiayu Yao, Soumya Ghosh, and Finale Doshi-Velez.
Latent projection bnns: Avoiding weight-space pathologies by learning latent representations of neural network weights. arXiv preprint arXiv:1811.07006, 2018.
[24] Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable Laplace approximation
for neural networks. In International Conference on Learning Representations (ICLR),
2018.
[25] Yunus Saatci and Andrew G Wilson. Bayesian GAN. In Advances in neural information
processing systems, pages 36223631, 2017.
[26] Matthias Seeger. Bayesian modelling in machine learning: A tutorial review. Technical
report, 2006.
[27] Shengyang Sun, Guodong Zhang, Jiaxin Shi, and Roger Grosse. Functional variational
bayesian neural networks. arXiv preprint arXiv:1903.05779, 2019.
[28] Christopher KI Williams and Carl Edward Rasmussen. Gaussian processes for machine
learning. the MIT Press, 2(3):4, 2006.
[29] Andrew Gordon Wilson. Covariance kernels for fast automatic pattern discovery and
extrapolation with Gaussian processes. PhD thesis, University of Cambridge, 2014.
[30] Andrew Gordon Wilson, Zhiting Hu, Ruslan Salakhutdinov, and Eric P Xing. Deep
kernel learning. In Artificial Intelligence and Statistics, pages 370378, 2016.
[31] Wanqian Yang, Lars Lorch, Moritz A Graule, Srivatsan Srinivasan, Anirudh Suresh,
Jiayu Yao, Melanie F Pradier, and Finale Doshi-Velez. Output-constrained bayesian
neural networks. arXiv preprint arXiv:1905.06287, 2019.
[32] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson.
Cyclical stochastic gradient MCMC for Bayesian deep learning. In International
Conference on Learning Representations, 2020.
[33] Konrad Zona, Krzysztof J Geras, and Kyunghyun Cho. Classifier-agnostic saliency map
extraction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33,
pages 1008710088, 2019.