Combining Markov Chain Monte Carlo and Stochastic Gradient Descent for in the sense of using SGD to do some cheap approximation to MCMC posterior sampling. Overviews in Ma, Chen, and Fox (2015) and Mandt, Hoffman, and Blei (2017). A lot of probabilistic neural nets leverage this idea.

A related idea is estimating gradients of parameters by Monte Carlo; there is nothing necessarily Bayesian about that *per se*; in that case we are doing a noisy estimate of a deterministic quantity.
In *this* setting we are interested in the noise itself.

I have a vague memory that this argument is leveraged in Neal (1996)? Should check. For sure the version in Mandt, Hoffman, and Blei (2017) is a highly developed and modern take. Basically, they analyse the distribution near convergence as an autoregressive process:

Stochastic Gradient Descent with a constant learning rate (constant SGD) simulates a Markov chain with a stationary distribution. With this perspective, we derive several new results.

- We show that constant SGD can be used as an approximate Bayesian posterior inference algorithm. Specifically, we show how to adjust the tuning parameters of constant SGD to best match the stationary distribution to a posterior, minimizing the Kullback-Leibler divergence between these two distributions.
- We demonstrate that constant SGD gives rise to a new variational EM algorithm that optimizes hyperparameters in complex probabilistic models.
- We also propose SGD with momentum for sampling and show how to adjust the damping coefficient accordingly.
- We analyze MCMC algorithms. For Langevin Dynamics and Stochastic Gradient Fisher Scoring, we quantify the approximation errors due to finite learning rates. Finally,
- we use the stochastic process perspective to give a short proof of why Polyak averaging is optimal. Based on this idea, we propose a scalable approximate MCMC algorithm, the Averaged Stochastic Gradient Sampler.

The article is rather beautiful. Importantly they leverage the assumption that we are sampling from approximately (log-)quadratic posterior modes, which means that we should be suspicious of the method when

- The posterior is not quadratic, i.e. the distribution is not well approximated by a Gaussian at the mode, and
- The same for the tails. If there are low-probability but high importance posterior configurations such that they are not Gaussian in the tails, we should be skeptical that they will be sampled well; I have an intuition that this is a more stringent requirement, but TBH I am not sure of the exact relationship of these two conditions.

The models leverage gradient flow, which is a continuous limit of stochastic gradient descent.

## Stochastic Weight Averaging

A popular recent development is the Stochastic Weight Averaging family of methods (Izmailov et al. 2018, 2020; Maddox et al. 2019; Wilson and Izmailov 2020). See Andrew G Wilsonβs web page for a brief description of the sub methods, since he seems to have been involved in all of them.

## Stochastic Gradient Langevin MCMC

βa Markov Chain reminiscent of noisy gradient descentβ (Welling and Teh 2011) extending vanilla Langevin dynamics.

## Stein Variational GD

Not quite what I mean, but related. An ensemble method. See Stein VGD.

## SG Hamiltonian Monte Carlo

## SG thermostats

- Some kind of variance control using auxiliary variables?

## SG Fisher scoring

Ahn, Korattikara, and Welling (2012)

## Incoming

## References

*Proceedings of the 29th International Coference on International Conference on Machine Learning*, 1771β78. ICMLβ12. Madison, WI, USA: Omnipress.

*Proceedings of the 39th International Conference on Machine Learning*, 414β34. PMLR.

*Proceedings of the 32nd International Conference on Neural Information Processing Systems*, 8278β88. NIPSβ18. Red Hook, NY, USA: Curran Associates Inc.

*2018 Information Theory and Applications Workshop (ITA)*, 1β10.

*Proceedings of the 31st International Conference on Machine Learning*, 1683β91. Beijing, China: PMLR.

*Proceedings of the 32nd International Conference on Neural Information Processing Systems*, 9187β97. NIPSβ18. Red Hook, NY, USA: Curran Associates Inc.

*Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2*, 3203β11. NIPSβ14. Cambridge, MA, USA: MIT Press.

*arXiv:1605.01559 [Math, Stat]*, May.

*arXiv:2105.04504 [Cs, Stat]*, May.

*arXiv:1812.00793 [Cs, Math, Stat]*, September.

*Journal of the Royal Statistical Society: Series B (Statistical Methodology)*73 (2): 123β214.

*Journal of the Royal Statistical Society: Series B (Methodological)*56 (4): 549β81.

*arXiv:1903.12322 [Cs, Stat]*, March.

*Proceedings of The 35th Uncertainty in Artificial Intelligence Conference*, 1169β79. PMLR.

*Advances In Neural Information Processing Systems*.

*Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2*, 2917β25. NIPSβ15. Cambridge, MA, USA: MIT Press.

*Proceedings of the 19th International Conference on Artificial Intelligence and Statistics*, 1070β77. arXiv.

*JMLR*, April.

*arXiv:1610.00781 [Math, Stat]*, October.

*Advances in Neural Information Processing Systems*. Vol. 32. Red Hook, NY, USA: Curran Associates, Inc.

*Nuclear Physics B*180 (3): 378β84.

*Statistics & Probability Letters*182 (March): 109321.

*Advances in Neural Information Processing Systems*. Vol. 28. NIPSβ15. Curran Associates, Inc.

*Proceedings of the 28th International Conference on International Conference on Machine Learning*, 681β88. ICMLβ11. Madison, WI, USA: Omnipress.

*Proceedings of the 37th International Conference on Machine Learning*, 10248β59. PMLR.

*Statistics & Probability Letters*91 (Supplement C): 14β19.

## No comments yet. Why not leave one?