Deep learning does not replace Bayesian modeling: Comparing research use via citation counting

Abstract

One could be excused for assuming that deep learning had or will soon usurp all credible work in reasoning, artificial intelligence, and statistics, but like most “meme” class broad generalizations the concept does not hold up to scrutiny. Memes do not generally matter since the experts will always know better; but in the case of Bayesian software like Stan and PyMC3, even their developers and advocates bemoan the apparent dominance of deep learning as manifested in popular culture, breathtaking performance, and most problematically from funding agency peer review that impacts our ability to further advance the field. The facts, however, do not support the assumed dominance of deep learning in science upon closer examination. This letter simply makes the argument by the crudest of possible metrics, citation count, that once the discipline of Computer Science is subtracted, Bayesian software accounts for nearly a third of research citations. Stan and PyMC3 dominate some fields, PyTorch, Keras, and TensorFlow dominate others with lot of variations in between. Bayesian and deep‐learning approaches are related but very different technologies in goals, implementation, and applicability with little actual overlap‐‐so this is not a surprise. For example, deep learning cannot bring the explainability of applied math/statistics and Bayesian methods do not scale to deep‐learning data sets. While deep‐learning behemoths like Facebook and Google use and support Bayesian efforts, the Bayesian packages scientists actually use are academic/volunteer efforts punching far above their weight class, and they need financial support. It would behoove funders to fully understand the impact and role of Bayesian methods in resource allocation.

DATA AVAILABILITY STATEMENT

All data/software required to re-run and create the paper are provided in a repository at: Raw results from queries are not included.

References

Carpenter B, Gelman A, Homan MD, et al. Stan: a probabilistic programming language. Grantee Submission. J Stat Softw. 2017;76(1):132.

Homan MD, Gelman A. The No‐U‐Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. J Mach Learn Res. 2014;15(1):1593‐1623.

Abadi M, Barham P, Chen J et al. Tensorow: a system for large‐scale machine learning. 2016;265‐283.

Lieu M. Hierarchical modelling of galaxy clusters for cosmology. StanCon Helsinki. 2018;2018.

Stan Development Team. Stan. 2012.

Salvatier J, Wiecki TV, Fonnesbeck C. Probabilistic programming in Python using PyMC3. PeerJ Comput Sci. 2016;2:e55.

Brkner PC. brms: an R package for Bayesian multilevel models using Stan. J Stat Softw. 2017;80(1): [eLocator: 128]. doi: [DOI: https://dx.doi.org/10.18637/jss.v080.i01]

Xie Y, Dervieux C, Riederer E. R Markdown Cookbook. Boca Raton, FL: Chapman and Hall/CRC; 2020.

Xie Y. Bookdown: Authoring Books and Technical Documents with R Markdown. Boca Raton, FL: Chapman and Hall/CRC; 2016.

RStudio Team. RStudio: Integrated Development Environment for R. Boston, MA: RStudio, PBC; 2020.

Zhu H. kableExtra: Construct Complex Table with “kable” and Pipe Syntax. 2021. R Package Version 1.3.4.

Wickham H, Seidel D. Scales: Scale Functions for Visualization; 2020 R Package Version 1.1.1.

Slowikowski K. ggrepel: Automatically Position Non‐Overlapping Text Labels with “ggplot2”. 2021. R Package Version 0.9.1.

Ooms J. The jsonlite Package: A Practical and Consistent Mapping Between JSON Data and R Objects. arXiv:1403.2805 [stat.CO]; 2014.

Wickham H. httr: Tools for Working with URLs and HTTP; 2020 R Package Version 1.4.2.

Wickham H. stringr: Simple, Consistent Wrappers for Common String Operations. 2019. R Package Version 1.4.0.

FitzJohn R. redux: R Bindings to “hiredis”. 2018. R package version 1.1.0.

Redis Development Team. Redis. 2020.