Stochastic optimization, uses noisy (possibly approximate) 1st-order gradient information to find the argument which minimises

\[ x^*=\operatorname{argmin}_{\mathbf{x}} f(x) \]

for some an objective function \(f:\mathbb{R}^n\to\mathbb{R}\).

That this works with little fuss in very high dimensions is a major pillar of deep learning.

The original version, given in terms of root finding, is (Herbert Robbins and Monro 1951) who later generalised analysis in (H. Robbins and Siegmund 1971), using martingale arguments to analyze convergence. There is some historical context in (Lai 2003) which puts it all in context. That article was written before the current craze for SGD in deep learning; after 2013 or so the problem is rather that there is so much information on the method that the challenge becomes sifting out the AI hype from the useful.

I recommend Francis Bach’s Sum of geometric series trick as an introduction to showing things advanced things about SGD using elementary tools.

🏗

Zeyuan Allen-Zhu : Faster Than SGD 1: Variance Reduction:

SGD is well-known for large-scale optimization. In my mind, there are two (and only two) fundamental improvements since the original introduction of SGD: (1) variance reduction, and (2) acceleration. In this post I’d love to conduct a survey regarding (1),

Francesco Orabana on how to prove SGD converges

to balance the universe of first-order methods, I decided to show how to easily prove the convergence of the iterates in SGD, even in unbounded domains.

## Variance-reduced

TBD

## Normalized

Zhiyuan Li and Sanjeev Arora argue:

You may remember our previous blog post showing that it is possible to do state-of-the-art deep learning with learning rate that increases exponentially during training. It was meant to be a dramatic illustration that what we learned in optimization classes and books isn’t always a good fit for modern deep learning, specifically,

normalized nets, which is our term for nets that use any one of popular normalization schemes,e.g. BatchNorm (BN), GroupNorm (GN), WeightNorm (WN). Today’s post (based upon our paper with Kaifeng Lyu at NeurIPS20) identifies other surprising incompatibilities between normalized nets and traditional analyses.

## Sundry Hacks

…

Yellowfin an automatic SGD momentum tuner

Mini-batch and stochastic methods for minimising loss when you have a lot of data, or a lot of parameters, and using it all at once is silly, or when you want to iteratively improve your solution as data comes in, and you have access to a gradient for your loss, ideally automatically calculated. It’s not clear at all that it should work, except by collating all your data and optimising offline, except that much of modern machine learning shows that it does.

Sometimes this apparently stupid trick it might even be fast for small-dimensional cases, so you may as well try.

Technically, “online” optimisation in bandit/RL problems might imply that you have to “minimise regret online”, which has a slightly different meaning and, e.g. involves seeing each training only as it arrives along some notional arrow of time, yet wishing to make the “best” decision at the next time, and possibly choosing your next experiment in order to trade-off exploration versus exploitation etc.

In SGD you can see your data as often as you want and in whatever order, but you only look at a bit at a time. Usually the data is given and predictions make no difference to what information is available to you.

Some of the same technology pops up in each of these notions of online optimisation, but I am really thinking about SGD here.

There are many more permutations and variations used in practice.

## References

*Advances in Neural Information Processing Systems (NIPS)*, –. Spain. http://hal.archives-ouvertes.fr/hal-00608041.

*Séminaire de Probabilités de Strasbourg*, 33:1–68. Lecture Notes in Math. Berlin: Springer - Lecture Notes in Mathematics. https://doi.org/10.1007/BFb0096509.

*Electronic Journal of Statistics*9 (2): 2058–75. https://doi.org/10.1214/15-EJS1071.

*Proceedings of Neuro-Nîmes 91*. Nimes, France: EC2. http://leon.bottou.org/papers/bottou-91c.

*Online Learning and Neural Networks*, edited by David Saad, 17:142. Cambridge, UK: Cambridge University Press. http://leon.bottou.org/publications/pdf/online-1998.pdf.

*Proceedings of the 19th International Conference on Computational Statistics (COMPSTAT’2010)*, 177–86. Paris, France: Springer. http://leon.bottou.org/papers/bottou-2010.

*Advances in Neural Information Processing Systems*, edited by J. C. Platt, D. Koller, Y. Singer, and S. Roweis, 20:161–68. NIPS Foundation (http://books.nips.cc). http://leon.bottou.org/papers/bottou-bousquet-2008.

*Advances in Neural Information Processing Systems 16*, edited by Sebastian Thrun, Lawrence Saul, and Bernhard Schölkopf. Cambridge, MA: MIT Press. http://leon.bottou.org/papers/bottou-lecun-2004.

*Convex Optimization: Algorithms and Complexity*. Vol. 8. Foundations and Trends in Machine Learning. Now Publishers. https://doi.org/10.1561/2200000050.

*IEEE Signal Processing Magazine*31 (5): 32–43. https://doi.org/10.1109/MSP.2014.2329397.

*Mathematical Programming*134 (1): 71–99. https://doi.org/10.1007/s10107-012-0569-0.

*Journal of Machine Learning Research*12 (July): 2121–59. http://www.jmlr.org/papers/v12/duchi11a.html.

*SIAM Journal on Scientific Computing*34 (3): A1380–1405. https://doi.org/10.1137/110830629.

*SIAM Journal on Optimization*23 (4): 2341–68. https://doi.org/10.1137/120880811.

*Advances in Neural Information Processing Systems 28*, edited by C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, 1594–1602. Curran Associates, Inc. http://papers.nips.cc/paper/5718-beyond-convexity-stochastic-quasi-convex-optimization.pdf.

*Stochastic Processes and Their Applications*2 (4): 359–70. https://doi.org/10.1016/0304-4149(74)90004-0.

*Advances in Neural Information Processing Systems*, 781–89. Curran Associates, Inc. http://papers.nips.cc/paper/3817-accelerated-gradient-methods-for-stochastic-optimization-and-online-learning.

*IEEE Transactions on Signal Processing*62 (4): 868–82. https://doi.org/10.1109/TSP.2013.2291221.

*Proceeding of ICLR*. http://arxiv.org/abs/1412.6980.

*The Annals of Statistics*31 (2): 391–406. https://doi.org/10.1214/aos/1051027873.

*Advances in Neural Information Processing Systems*, 2283–91. http://papers.nips.cc/paper/5129-stochastic-majorization-minimization-algorithms-for-large-scale-optimization.

*JMLR*, April. http://arxiv.org/abs/1704.04289.

*Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining*, 1222–30. KDD ’13. New York, NY, USA: ACM. https://doi.org/10.1145/2487575.2488200.

*SIAM Journal on Control and Optimization*30 (4): 838–55. https://doi.org/10.1137/0330046.

*PMLR*, 1603:314–23. http://proceedings.mlr.press/v48/reddi16.html.

*Optimizing Methods in Statistics*, edited by Jagdish S. Rustagi, 233–57. Academic Press. https://doi.org/10.1016/B978-0-12-604550-5.50015-8.

*The Annals of Mathematical Statistics*22 (3, 3): 400–407. https://doi.org/10.1214/aoms/1177729586.

*Advances in Neural Information Processing Systems 29*, edited by D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, 901–1. Curran Associates, Inc. http://papers.nips.cc/paper/6114-weight-normalization-a-simple-reparameterization-to-accelerate-training-of-deep-neural-networks.pdf.

*Journal of Machine Learning Research*12 (July): 1865–92. http://www.machinelearning.org/archive/icml2009/papers/262.pdf.

*IEEE Transactions on Automatic Control*45 (10): 1839–53. https://doi.org/10.1109/TAC.2000.880982.

*Proceedings of the 23rd International Conference on Machine Learning*.

*Advances in Neural Information Processing Systems 23*, edited by J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta, 2595–2603. Curran Associates, Inc. http://papers.nips.cc/paper/4006-parallelized-stochastic-gradient-descent.pdf.