Do you know Google Summer of Code? It’s a time of year when students can contribute to open-source software by developing and adding much needed functionalities to the open-source package of their choice. And Demetri Pananos did just that.

He did it in 2019 with PyMC3, for which he developed the API for ordinary differential equations. In this episode, he’ll tell us why and how he did that, what he learned from the experience, and what the strengths and weaknesses of the API are in his opinion.

Demetri is a Ph.D candidate in Biostatistics at Western University, in Ontario, Canada. His research interests surround machine learning and Bayesian statistics for personalized medicine. He earned his Master’s in Applied Mathematics from The University of Waterloo and is a firm believer in open science, interdisciplinary collaboration, and reproducible research. 

Other than that, he loves plotting data and drinking IPA beer – well, who doesn’t?”

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Demetri on Twitter: https://twitter.com/PhDemetri
• Demetri on GitHub: https://github.com/Dpananos
• Demetri's website: https://dpananos.github.io/
• PyMC3, Probabilistic Programming in Python: https://docs.pymc.io/
• Chris Bishop, Pattern Recognition and Machine Learning: https://www.amazon.fr/Pattern-Recognition-Machine-Learning-Christopher/dp/0387310738
• Bayesian Data Analysis (Gelman, Carlin, Stern, Dunson, Vehtari, Rubin): http://www.stat.columbia.edu/~gelman/book/
• Parallel Plots: https://arviz-devs.github.io/arviz/generated/arviz.plot_parallel.html

I bet you already heard about hierarchical models, or multilevel models, or varying-effects models — yeah this type of models has a lot of names! Many people even turn to Bayesian tools to build _exactly_ these models. But what are they? How do you build and use a hierarchical model? What are the tricks and classical traps? And even more important: how do you _interpret_ a hierarchical model?

In this episode, Thomas Wiecki will come to the rescue and explain what multilevel models are, how to build them, what their powers are… but also why you should be very careful when building them…

Does the name Thomas Wiecki ring a bell? Probably because he’s the host and creator of the PyData Deep Dive Podcast, where he interviews open-source contributors from the Python and Data Science worlds! Thomas is also the VP of Data Science at Quantopian, a crowd-sourced quantitative investment firm that encourages people everywhere to write investment algorithms.

Finally, Thomas is a longtime Bayesian and core-developer of PyMC3, a fantastic python package to do probabilistic programming in Python. On his blog, he publishes tutorial articles and explores new ideas such as Bayesian Deep Learning. Caring a lot about open-source software sustainability, he puts all he’s up to on his Patreon page, that you’ll find in the show notes.

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Thomas’ series on Hierarchical Regression: https://twiecki.io/blog/2013/08/12/bayesian-glms-1/
• Non-centered Parametrization with PyMC3: https://twiecki.io/blog/2017/02/08/bayesian-hierchical-non-centered/
• Using Bayesian Decision Making: https://twiecki.io/blog/2019/01/14/supply_chain/
• PyMC3 - Probabilistic Programming in Python: https://docs.pymc.io/
• Symbolic PyMC: https://pymc-devs.github.io/symbolic-pymc/
• PyData Deep Dive Podcast: https://pydata-podcast.com
• Thomas on Twitter: https://twitter.com/twiecki?lang=en
• Thomas on Patreon: https://www.patreon.com/twiecki
• Thomas on GitHub: https://github.com/twiecki
• Alex’s Hierarchical Model of Elections in Paris: https://mybinder.org/v2/gh/AlexAndorra/pollsposition_models/master?urlpath=%2Fvoila%2Frender%2Fdistrict-level%2Fmunic_model_analysis.ipynb

How do you handle your MCMC samples once your Bayesian model fit properly? Which diagnostics do you check to see if there was a computational problem? And isn’t that nice when you have beautiful and reliable plots to complement your analysis and better understand your model?

I know what you think: plotting can be long and complicated in these cases. Well, not with ArviZ, a platform-agnostic package to do exploratory analysis of your Bayesian models. And in this episode, Ari Hartikainen will tell you why.

Ari is a data-scientist in geophysics and a researcher at the Department of Civil Engineering of Aalto University in Finland. He mainly works on geophysics, Bayesian statistics and visualization. 

Ari’s also a prolific open-source contributor, as he’s a core-developer of the popular Stan and ArviZ libraries. He’ll tell us how PyStan interacts with ArviZ, what he thinks ArviZ most useful features are, and which common difficulties he encounters with his models and data.

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Ari on GitHub: https://github.com/ahartikainen
• Ari on Twitter: https://twitter.com/a_hartikainen
• ArviZ -- Exploratory analysis of Bayesian models: https://arviz-devs.github.io/arviz/
• Introductory paper of ArviZ in The Journal of Open Source Software: https://www.researchgate.net/publication/330402908_ArviZ_a_unified_library_for_exploratory_analysis_of_Bayesian_models_in_Python
• Stan -- Statistical Modeling Platform: https://mc-stan.org/
• GPflow -- Gaussian processes in TensorFlow: https://www.gpflow.org/
• GPy -- Gaussian processes framework in Python: https://sheffieldml.github.io/GPy/

Have you always wondered what dark matter is? Can we even see it — let alone measure it? And what would discover it imply for our understanding of the Universe?

In this episode, we’ll take look at the cosmos with Maggie Lieu. She’ll tell us what research in astrophysics is made of, what model she worked on at the European Space Agency, and how Bayesian the world of space science is.

Maggie Lieu did her PhD in the Astronomy & Space Department of the University of Birmingham. She’s now a Research Fellow of Machine Learning & Cosmology at the University of Nottingham and is working on projects in preparation for Euclid, a space-based telescope whose goal is to map the dark Universe and help us learn about the nature of dark matter and dark energy.

In a nutshell, she tries to help us better understand the entire cosmos. Even more amazing, she uses the Stan library and applies Bayesian statistical methods to decipher her astronomical data! But Maggie is not just a Bayesian astrophysicist: she also loves photography and rock-climbing!

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Maggie's Website: https://maggielieu.com/
• Maggie's Google Scholar Page: https://scholar.google.co.uk/citations?user=ilfwfuUAAAAJ&hl=en
• Maggie on Twitter: https://twitter.com/Space_Mog
• Maggie on GitHub: https://github.com/MaggieLieu
• Maggie on YouTube: https://www.youtube.com/channel/UClO6TuRE6XLzbMBmQ_KY38A
• Stan -- Statistical Modeling Platform: https://mc-stan.org/
• Stan's YouTube Channel: https://www.youtube.com/channel/UCwgN5srGpBH4M-Zc2cAluOA

What is it like using Bayesian tools when you’re a software engineer or computer scientist? How do you apply these tools in the online ad industry? 

More generally, what is Bayesian thinking, philosophically? And is it really useful in every day life? Because, well you can’t fire up MCMC each time you need to make a quick decision under uncertainty… So how do you do that in practice, when you have at most a pen and paper?

In this episode, you’ll hear Max Sklar’s take on these questions. Max is a software engineer with a focus on machine learning and Bayesian inference. Now working at Foursquare’s innovation lab, he recently led the development of a causality model for Foursquare’s Ad Attribution product and taught a course on Bayesian Thinking at the Lviv Data Science Summer School.

Max is also an open-source enthusiast and a fellow podcaster – he’s the host of the Local Maximum podcast, where you can hear every week about the latest trends in AI, machine learning and technology from an engineering perspective.

Ow, and if you liked the movie « Her », with Joaquin Phoenix, well you’re in for a treat at the end of this episode…

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Local Maximum podcast website: https://www.localmaxradio.com
• Max on Twitter: https://twitter.com/maxsklar
• Bayesian linear models: https://github.com/maxsklar/BayesPy/tree/master/LinearModels
• Bayesian Dirichlet-Multinomial estimation: https://github.com/maxsklar/BayesPy/tree/master/DirichletEstimation
• Bayesian Thinking for Applied Machine Learning slides: https://docs.google.com/presentation/d/1eiceuvXlsoFKoHdqjF3qXBkyht7vR0YXQPG82ady-TU/edit?usp=sharing

You can’t study psychology up until your PhD and end-up doing very mathematical and computational data science at Google right? It’s too hard of a U-turn — some would even say it’s NUTS, just because they like bad puns… Well think again, because Junpeng Lao did just that!

Before doing data science at Google, Junpeng was a cognitive psychology researcher at the University of Fribourg, Switzerland. Working in Python, Matlab and occasionally in R, Junpeng is a prolific open-source contributor, particularly to the popular TensorFlow and PyMC3 libraries. He also maintains the PyMC Discourse on his free time, where he amazingly answers all kinds of various and very specific questions!

In this episode, he’ll tell you what the core characteristics of TensorFlow Probability are, and when you would use TFP instead of another probabilistic programming framework, like Stan or PyMC3. He’ll also explain why PyMC4 will be based on TensorFlow Probability itself, and what future contributions he has in mind for these two amazing libraries. Finally, Junpeng will share with you his workflow for debugging a model, or just for better understanding your models.

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show: 

• Junpeng's blog: https://junpenglao.xyz/
• Junpeng on Twitter: https://twitter.com/junpenglao
• Junpeng on GitHub: https://github.com/junpenglao
• Advanced Bayesian Modeling Tutorial: https://discourse.pymc.io/t/advance-bayesian-modelling-with-pymc3/1439
• Stan Devs' Prior Choice Recommendations: https://github.com/stan-dev/stan/wiki/Prior-Choice-Recommendations
• PyMC Discourse: https://discourse.pymc.io/
• PyMC3 - Probabilistic Programming in Python: https://docs.pymc.io/
• Tensor Flow Probability: https://www.tensorflow.org/probability/

If you’re there, it’s probably because you’re interested in Bayesian inference, right? But don’t you feel lost sometimes when building a model? Or you ask yourself why what you’re trying to do is so damn hard… and you conclude that YOU are the problem, that YOU must be doing something wrong!

Well, rest assured, as you’ll hear from Michael Betancourt himself: it’s hard for everybody! That’s why over the years he developed and tries to popularize what he calls a « principled Bayesian workflow » — in a nutshell, think about what could have generated your data; and always question default settings!

With that workflow, you’ll probably feel less alone when modeling, but expect to fail often. That’s ok — as Michael says: if you don’t fail, you don’t learn!

Who is Michael Betancourt you ask? He is a physicist and statistician, whose research focuses on the development of robust statistical workflows, computational tools, and pedagogical resources that help bridge the gap between statistical theory and scientific practice.

Michael works a lot on differential geometry and probability theory, and he often lives in high-dimensional spaces, where he meets with a good friend of his -- Hamiltonian Monte Carlo. Then, you won’t be surprised to learn that Michael is one of the core developers of the seminal probabilistic programming language Stan.

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Michael's upcoming course: https://events.eventzilla.net/e/introduction-to-bayesian-inference-with-stan-with-michael-betancourt-2138756860
• Michael's website (the “Writing” page collects the case studies and pedagogical material, and the “Speaking” page links to the recorded talks): https://betanalpha.github.io/
• Support Michael's work on Patreon: https://patreon.com/betanalpha
• Michael on Twitter: https://twitter.com/betanalpha
• Michael on GitHub: https://github.com/betanalpha
• Stan probabilistic programming langage: https://mc-stan.org/

I have two questions for you: Are you a self-learner? Then how do you stay up to date? What should you focus on if you’re a beginner, or if you’re more advanced?

And here is my second question: Are you working in biomedicine? And if you do, are you using Bayesian tools? Then how do you get your co-workers more used to posterior distributions than p-values? In other words, how do you change behaviors in a large organization?

In this episode, Eric Ma will answer all these questions and even tell us his favorite modeling techniques, which problems he encountered with these models, and how he solved them. He’ll also share with us the software-engineering workflow he uses at Novartis to share his work with colleagues.

Eric is a data scientist at the Novartis Institutes for Biomedical Research, where he focuses on Bayesian statistical methods to make medicines for patients. Eric is also a prolific open source developer: he led the development of pyjanitor, an API for cleaning data in Python, and nxviz, a visualization package for NetworkX. He also contributes to PyMC3, matplotlib and bokeh.

This is « Learning Bayesian Statistics », episode 5, recorded October 21, 2019.

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Links from the show:

• Eric's website: https://ericmjl.github.io/
• Eric on Twitter: https://twitter.com/ericmjl
• Bayesian analysis recipes: https://github.com/ericmjl/bayesian-analysis-recipes
• Bayesian deep learning demystified: https://github.com/ericmjl/bayesian-deep-learning-demystified
• Causality repo: https://github.com/ericmjl/causality
• Pyjanitor - Convenient data cleaning routines for repetitive tasks: https://pyjanitor.readthedocs.io/
• PyMC3 - Probabilistic Programming in Python: https://docs.pymc.io/
• Panel - A high-level app and dashboarding solution for Python: https://panel.pyviz.org/
• Nxviz - Visualization Package for NetworkX: https://nxviz.readthedocs.io/en/latest/

What do neurodegenerative diseases, gerrymandering and ecological inference all have in common? Well, they can all be studied with Bayesian methods — and that’s exactly what Karin Knudson is doing.

In this episode, Karin will share with us the vital and essential work she does to understand aspects of neurodegenerative diseases. She’ll also tell us more about computational neuroscience and Dirichlet processes — what they are, what they do, and when you should use them.

Karin did her doctorate in mathematics, with a focus on compressive sensing and computational neuroscience at the University of Texas at Austin. Her doctoral work included applying hierarchical Dirichlet processes in the setting of neural data and focused on one-bit compressive sensing and spike-sorting.

Formerly the chair of the math and computer science department of Phillips Academy Andover, she started a postdoc at Mass General Hospital and Harvard Medical in Fall 2019. Most importantly, rock climbing and hiking have no secrets for her!

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ ! 

Links from the show, personally curated by Karin Knudson:

• Karin on Twitter: https://twitter.com/karinknudson
• Spike train entropy-rate estimation using hierarchical Dirichlet process priors (Knudson and Pillow): https://pillowlab.princeton.edu/pubs/abs_Knudson_HDPentropy_NIPS13.html
• Fighting Gerrymandering with PyMC3, PyCon 2018, Colin Carroll and Karin Knudson: https://www.youtube.com/watch?v=G9I5ZnkWR0A
• Expository resources on Dirichlet Processes: Chapter 23 of Bayesian Data Analysis (Gelman et al.) and http://www.gatsby.ucl.ac.uk/~ywteh/research/npbayes/dp.pdf
• Hierarchical Dirichlet Processes (introduced the HDP and included applications in topic modeling and for working with time-series data and Hidden Markov Models): https://www.stat.berkeley.edu/~aldous/206-Exch/Papers/hierarchical_dirichlet.pdf
• A Sticky HDP-HMM with applications to speaker diarization (a nice example of how the HDP can be used with HMM, in this case cleverly adapted so that states have more persistence): https://arxiv.org/abs/0905.2592
• If you want to get deeper into the weeds and also get a sense of the history: Dirichlet Processes with Applications to Bayesian Nonparametric Problems (https://projecteuclid.org/euclid.aos/1176342871) and A Bayesian Analysis of Some Nonparametric Problems (https://projecteuclid.org/euclid.aos/1176342360)

How can you use Bayesian tools and optimize your models in industry? What are the best ways to communicate and visualize your models with non-technical and executive people? And what are the most common pitfalls?

In this episode, Colin Carroll will tell us how he did all that in finance and the airline industry. He’ll also share with us what the future of probabilistic programming looks like to him.

You already heard from Colin two weeks ago — so, if you didn’t catch this episode, go back in your feed’s history and enjoy the first part! 

As a reminder, Colin is a machine learning researcher and software engineer who’s notably worked on modeling risk in the airline industry and building NLP-powered search infrastructure for finance. He’s also an active contributor to open source, particularly to the popular PyMC3 and ArviZ libraries.

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/!

Links from the show:

•     Colin's blog: https://colindcarroll.com/
•     Gelman’s putting model in PyMC3: https://github.com/pymc-devs/pymc3/blob/master/docs/source/notebooks/putting_workflow.ipynb
•     Matthew Kay’s quantile dotplots: https://github.com/mjskay/when-ish-is-my-bus/blob/master/quantile-dotplots.md
•     Jax, Composable transformations of Python+NumPy programs: https://github.com/google/jax
•     NumPyro, Probabilistic programming with NumPy: https://github.com/pyro-ppl/numpyro
•     Pyro, Deep Universal Probabilistic Programming: https://pyro.ai/
•     Rainier, Bayesian inference in Scala: https://github.com/stripe/rainier

---

Send in a voice message: https://anchor.fm/learn-bayes-stats/message