Despite the promise of big data, inferences are often limited not by the size of data but rather by its systematic structure. Only by carefully modeling this structure can we take fully advantage of the data -- big data must be complemented with big models and the algorithms that can fit them. Stan is a platform for facilitating this modeling, providing an expressive modeling language for specifying bespoke models and implementing state-of-the-art algorithms to draw subsequent Bayesian inferences.
These modules present a modern perspective on Bayesian modeling, beginning with a principled Bayesian workflow and then progressing to in depth reviews of popular modeling techniques. The course will emphasize interactive exercises run through RStan, the R interface to Stan, and PyStan, the Python interface to Stan.
The course is aimed at current Stan users and will assume familiarity with the basics of calculus, linear algebra, probability theory, probabilistic modeling and statistical inference, and Stan.
Attendees are strongly encouraged to review
Conditional Probability Theory
Probability Theory on Product Spaces
Sampling and The Monte Carlo Method
Common Families of Probability Density Functions
Probabilistic Modeling and Inference
Markov chain Monte Carlo
Introduction to Stan
In order to participate in the exercises attendees must have a computer with RStan 2.19 and at least R 4.0 (https://cran.r-project.org/web/packages/rstan/index.html) or PyStan 2.19 and at least Python 3.7 (https://pystan2.readthedocs.io/en/latest/) installed. Please verify that you can run the 8 schools model as discussed in https://github.com/stan-dev/rstan/wiki/RStan-Getting-Started or https://pystan2.readthedocs.io/en/latest/getting_started.html and report any installation issues at http://discourse.mc-stan.org as early as possible. While the exercises cannot be run directly with other Stan interfaces such as CmdStanR and CmdStanPy they are relatively straightforward to translate.
Right before this course Jumping Rivers will also be offering introductory courses in RStan and PyStan that will provide an excellent review of much of the prerequisite materiall. I will add links to the courses once they're available.
Each module, except for the first, consists of a live lecture followed by exercises for attendees to complete individually before a final live review session. The modules are offered in parallel morning and afternoon sessions for scheduling flexibilty; both morning and afternoon sessions cover the same material.
The lecture and review will take place on Whereby which does not require any registration from attendees. Both will be recorded and and the recording will be made available to course attendees soon after each concludes. By participating you consent to having any video and audio you transmit recorded and shared with the rest of the course attendees. Not sharing your video and audio will not prevent you from viewing the sessions or asking questions through the interactive chat.
In between the lecture and review session attendees will be able to discuss the lecture and the exercises with each other on Discord. You should be able to access the course-specific Discord servers without any registration.
Module 1: Probabilistic and Generative Modeling
Module 1 discusses probabilistic modeling with an emphasis on generative models that capture the particular details of an assumed data generating process. The first lecture reviews the conceptual foundations of probabilistic modeling and statistical inference while the second lecture focuses on the powerful practical benefits of generative modeling.
First Lecture: Mon Feb 21 10:00 EST - 12:00 EST
Second Lecture: Thursday Feb 24 10:00 EST - 12:00 EST
First Lecture: Mon Feb 21 15:00 EST - 17:00 EST
Second Lecture: Thursday Feb 24 15:00 EST - 17:00 EST
Module 2: Identifiability and Degeneracy
The second module introduces the quantitative concept of the identifiability of probabilistic models and the more qualitative, but more practically relevant, concept of degeneracy of statistical inferences. We will review strategies for not only identifying and investigating degenerate behavior but also managing the pathological consequences of that behavior.
Lecture: Mon Feb 28 10:00 EST - 12:00 EST
Review Session: Thursday Mar 3 10:00 EST - 11:30 EST
Lecture: Mon Feb 28 15:00 EST - 17:00 EST
Review Session: Thursday Mar 3 15:00 EST - 16:30 EST
Module 3: Principled Bayesian Workflow
In this module we review a principled Bayesian workflow that guides the development of statistical models suited to the particular details of a given application. The workflow integrates the development of prior models, computational calibration, inferential calibration, and model critique and model updating.
Lecture: Mon Mar 7 10:00 EST - 12:00 EST
Review Session: Thursday Mar 10 10:00 EST - 11:30 EST
Lecture: Mon Mar 7 15:00 EST - 17:00 EST
Review Session: Thursday Mar 10 15:00 EST - 16:30 EST
Module 4: Regression Modeling
This module presents linear and general linear regression techniques from a modeling perspective, using that context to motivate robust implementations. We will especially emphasize principled prior modeling strategies for linear, log, and logistic regression models.
Lecture: Mon Mar 14 10:00 EST - 12:00 EDT
Review Session: Thursday Mar 17 10:00 EDT - 11:30 EDT
Lecture: Mon Mar 14 15:00 EST - 17:00 EDT
Review Session: Thursday Mar 17 15:00 EDT - 16:30 EDT
Module 5: Hierarchical Modeling
Module 5 introduces exchangeability and hierarchical models with a strong focus on the inherent identifiability issues and their computational consequences, as well as strategies for moderating this issues.
Lecture: Mon Mar 21 10:00 EST - 12:00 EDT
Review Session: Thursday Mar 24 10:00 EDT - 11:30 EDT
Lecture: Mon Mar 21 15:00 EST - 17:00 EDT
Review Session: Thursday Mar 24 15:00 EDT - 16:30 EDT
Module 6: Factor Modeling
This module introduces conditional exchangeability, marginal exchangeability, and observed factor modeling (also known as multilevel or random effects modeling) with a focus on efficient implementations.
Completion of Module 5 is highly recommended.
Lecture: Mon Mar 28 10:00 EST - 12:00 EDT
Review Session: Thursday Mar 31 10:00 EDT - 11:30 EDT
Lecture: Mon Mar 28 15:00 EST - 17:00 EDT
Review Session: Thursday Mar 31 15:00 EDT - 16:30 EDT
Module 7: Gaussian Process Modeling
The seventh module introduces Gaussian processes as a statistical modeling technique, motivating principled prior models that avoid pathological behavior.
Lecture: Mon Apr 4 10:00 EDT - 12:00 EDT
Review Session: Thursday Apr 7 10:00 EDT - 11:30 EDT
Lecture: Mon Apr 4 15:00 EDT - 17:00 EDT
Review Session: Thursday Apr 7 15:00 EDT - 16:30 EDT
Module 8: Sparsity Modeling
The final module reviews the concept of sparsity in Bayesian inference and prior modeling techniques that encourage sparse inferences.
Completion of Module 5 is highly recommended.
Lecture: Mon Apr 11 10:00 EDT - 12:00 EDT
Review Session: Thursday Apr 14 10:00 EDT - 11:30 EDT
Lecture: Mon Apr 11 15:00 EDT - 17:00 EDT
Review Session: Thursday Apr 14 15:00 EDT - 16:30 EDT
Cancellations will be considered only in the event of emergencies. Those not able to attend the modules due to unexpected scheduling conflicts will be able to follow along with the recordings and Discord discussion groups.
“We had a brilliant 3-day course at trivago with Michael Betancourt! The first day was filled with a very strong theoretical foundation for statistical modelling/decision making, followed by a crash course on MCMC and finished off with practical examples on how to diagnose healthy model fitting. In the 2nd and 3rd days we learned about many different types of hierarchical/multi-level models and spent most of the time practicing how to actually create and fit these models in Stan.
Michael is both a very engaging teacher, a very knowledgeable statistical modeller and, of course, a Stan master. This course has opened up new ways for us at trivago to gain better insights from our data through Stan models that fit our needs.”
-Data Scientists in the Automated Bidding Team, trivago
“The 1-day training course provided a great introduction to Bayesian models and their implementation in the Stan language. The practical focus really helped jumpstart our transition to Bayesian methods, and the slides, recorded lecture, and exercises also provide a great resource for new group members.”
-Stanley Lazic, Associate Director in Statistics and Machine Learning, AstraZeneca
“Stan is the cream of the crop platform for doing Bayesian analysis and is particularly appealing because of its open source nature. The programming language and algorithms are well designed and thought out. With that said, Stan has a very steep learning curve requiring lots of hours to get up to speed on your own. I have been to two training courses taught by Dr. Michael Betancourt and took an opportunity to have some consulting time. These sessions have proven invaluable to improve my use of Stan, increased my learning and usage rate, and informed me how to diagnose and detect issues that will inevitable will arise.”
-Robert Johnson, Corporate R&D, Procter & Gamble
“The workshop at MIT led by Michael Betancourt was a fun and very useful introduction to Stan. Mike worked with us to customize the lectures to our interests, he presented the material in an engaging and accessible way, and the physicists who attended, many of whom had never used Stan before, left with the resources to begin developing our own analyses. Mike’s background in physics makes him an especially effective teacher for scientists. The coding exercises were thoughtfully developed to progress in complexity and were well-integrated into the course; having such useful exercises was critical for participants to successfully internalize the concepts presented in the lectures.”
-Elizabeth Worcester, Associate Physicist, Department of Physics, Brookhaven National Laboratory