I started a YouTube channel! The videos will broadly fall into these categories:
- Academia as an early career researcher
- Technical data science and ML
- Applied R and Python
In my first video, I talk about my background and my plans for the channel:
@tcarpenter216
Machine learning scientist here.
I use the same trick to get rich with poker ♣️♥️♠️♦️
1. Play 649,739 hands privately at home.
2. Go to the casino‘s poker table
3. Go all-in in the first round
4. Get a guaranteed royal flush (odds 649,739 : 1)
Casinos hate this easy trick! 🔥
@lies_and_stats
@tcarpenter216
Because I want you all to sign up for my free newsletter which is really just a sales funnel into my 1,399$ masterclass.
1. Many consider brms to be one of the best pieces of statistical software out there.
2. Being Paul's PhD student for the last 3 years, I know for sure that his writing skills are next-level.
Combining 1 and 2, this book is bound to be great.
For the nerds: p(great | 1, 2) =💯
Drafts of the first chapters of my brms Book: Applied Bayesian Regression Modelling Using R and Stan are online: Check it out and let me know what you think!
Psychologists make great Data Scientists! 🚀
I hold two Master’s degrees:
→ one in Data and Computer Science📊🤖
→ one in Psychology🧠🗣️
Studying Psychology has made me a better data scientist, and here's why you should hire Psychologists for your data science team, too 🧵↓
I am excited to start a new chapter: Today is my first day as a PhD candidate in
@paulbuerkner
's research group for Bayesian Statistics at
@SimTechStuttga2
.
I'm extremely grateful for this amazing opportunity and full of joy & curiosity about the future in this awesome team!
Do you want to create a personal website but don't know where to start?
My latest blog post is perfect for you.🚀
Learn how to create a website with
#Quarto
and my simple template to get started in no time!
👇 Check it out and create your website now.
Our paper on consistency models for simulation-based inference has been accepted at
#NeurIPS2024
!
Thanks to a fantastic team Valentin Pratz (co-lead), Ullrich Köthe,
@paulbuerkner
,
@StefanRadev13
📄 Preprint:
🔁 Updates, code, and thread follow soon™
We have pre-released the {ggsimplex} R package. It is a ggplot extension for point and density plots in the probability 2-simplex (triangle). Useful for visualizing Posterior Model Probabilities in
#Bayesian
#Statistics
.
👉 Find out more in the blog post:
@ZetaOf1
@Julius_Ktxt
I can somewhat understand it from a mathematician‘s perspective. Assignment and equality aren’t the same:
x = x+1 ⁉️
vs.
x <- x+1 ✅
Our new short paper on Amortized Bayesian Workflow is out!✨
We developed an adaptive workflow that combines the speed of amortized inference with the reliability of MCMC on thousands of datasets.
🔗Link:
The whole is more than the sum of its parts 🧵👇
@FanLiDuke
And of course the infamous "I‘m humbled by xx" (also awards)
Which makes no sense at all because winning an award literally means you best everyone else how can that humble you lol
After announcing my new YouTube channel, it got >150 subscribers and >700 views.
These numbers might seem small for all the large accounts out there, but it means a lot to me.
I didn’t expect such a resonance. Thank you!
I’m already working on the next video, stay tuned 👀
Ever wondered whether your extreme evidence for a model is reproducible in a replication study?
We propose a method to predict the posterior model probabilities on new data with a novel framework: Meta-Uncertainty.
The best thing: It's applicable to virtually any analysis!
Posterior model probabilities are measures of uncertainty, but are also uncertain themselves. In our new preprint,
@MarvinSchmittML
@StefanRadev13
and I take a deep dive: We introduce Meta-Uncertainty and explore ties to overconfidence and reproducibility:
Thank you so much for the shoutout!
The brand-new BayesFlow version based on Keras 3 is available on the dev branch. We'll merge to main when we're feature-complete.
Questions, feedback, and contributions are always welcome!
Here is the Discourse forum:
BayesFlow is a neat library for simulation-based Amortized Bayesian Inference with deep learning.
It is based on Keras 3, which unlocks compatibility with all major frameworks -- JAX, PyTorch, TensorFlow, and NumPy.
🚨 BayesFlow can reliably detect Model Misspecification and Posterior Errors in Amortized Bayesian Inference 🚨
In a paper with
@paulbuerkner
, Ullrich Köthe, and
@StefanRadev13
(), we tackle a major problem in simulation-based inference w/ neural nets! 🧵👇
Working with Paul is the best thing that ever happened to me in my professional life. It's hard to overstate how grateful I am that he is my PhD supervisor.
If you want to apply to this position and have any informal questions: My DMs are open!
In my PhD, agile methods help me work more efficiently, achieve my research goals with greater ease, and make me feel less overwhelmed overall.
Here are 6 straightforward agile methods that I have found useful. Try them out for yourself and see the difference they can make. 🧵
All equations in my presentation (Quarto+revealjs) suddenly stopped rendering properly.
I just spent an hour debugging.
Turns out the train WiFi died and MathJax doesn't work offline.
AMA 🫠
As a scientist, your personal website is more than just a digital publication list.
It's a powerful tool to showcase your research, build your professional brand, and connect with peers around the globe.
Here's why every scientist should have a personal website:
I'll be presenting our paper "Meta-Uncertainty in Bayesian Model Comparison" at today's AISTATS poster session.
👉 Let's chat at poster
#125
this afternoon!
🌐 Project website:
#AISTATS2023
@aistats_conf
@aistats2023
This is the most exotic application of our BayesFlow library for amortized Bayesian inference so far:
🕊️ An agent-based model of birdsong learning on real data 🕊️
Authors:
@oporornis
,
@aalbina
, R Alexander Bentley, David Guerra,
@MasYoungblood
🔗 Link:
@PR0GRAMMERHUM0R
Nothing wrong here.
According to naming conventions, there’s a HelloWorld class with a constructor that takes a string argument.
In this line of code, an instance of HelloWorld is initialized with the argument "print".
Problem? 🥸
✨ I had a blast on the
@LearnBayesStats
podcast
🧪 We talked about my research on amortized Bayesian inference (=neural nets for fast Bayes)
💭 I'm passionate about this promising field, but many open questions remain on reliability
🎧 Tune in and let us hear your thoughts!
📢 Episode 107 is Now Available!
✨ In this episode, we dive deep into the world of amortized Bayesian inference with
@MarvinSchmittML
🎧 Tune in now to gain insights:
I am writing a guide on the essentials of creating a personal website. Aimed primarily at academics, but open to anyone familiar with a computer and comfortable with RStudio.
👉 What topics would you like to see covered in this tutorial blog post?
I am excited to announce the first publication of our new industry-academia collaboration.
Together with great co-authors
@tcarpenter216
@ZetaOf1
, we explore the impact of graduate education on subjective well-being.
Link to full pre-print (YouTube):
The sun is setting over Levi and
@bayescomp
has ended.
Thanks to all attendees for the great conversations and talks. And a special thank you to the organizers for making this awesome conference possible!
#BayesComp2023
On my way to the
@ELLISRobustML
workshop in beautiful Helsinki 🇫🇮☀️
I will present our paper on data-efficient amortized Bayesian inference with self-consistency losses (ICML 2024) in today‘s poster session.
1/2
We carry that to the extreme in our Meta-Uncertainty framework for Bayesian model comparison.
We simulate data from each competing model and perform full model comparisons on each simulated data set. This approaches the *sampling distribution* of model comparison results
I’m never getting tired of annoying my coauthors, course participants and consultees in suggesting to fit a statistical model to simulated data first, before you unleash it on experimental or field data. Happy to see these 3 recent articles in my favor. ⬇️
I am presenting a poster at
@bayescomp
today.
Come by and let's chat about why we have to think extra hard about model misspecification when we perform amortized Bayesian inference.
#BayesComp2023
Had a great time hosting a tooling session about amortized simulation-based inference at the ELLIS Doctoral Symposium.
So many amazing tooling sessions from inspiring colleagues. It was a pleasure working with all of you!
@ELLISforEurope
@FCAI_fi
#EDS2023
#ELLISPhD
How can we use deep neural networks for Bayesian posterior inference? ✨
What's the difference to a Bayesian Neural Network (BNN)? 👀
Tune in to episode
#107
of the Learning Bayesian Statistics podcast for more about Deep Learning ⚭ Bayes 🎧
Catch the full episode below 👇
Our paper on misspecification in amortized Bayesian inference was awarded with an Honorable Mention at the German Conference on Pattern Recognition 2023
@gcpr2023
.
Thanks to my great co-authors Paul Bürkner (
@paulbuerkner
), Ullrich Köthe & Stefan Radev (
@StefanRadev13
)!
Link👇
@eleafeit
Here’s a great paper from my group that disentangles the question:
TLDR: The approximation algorithm is one part of the model, the other ones are (1) the *probabilistic joint model* and (2) the data to fit.
I will give a talk about Reliable Amortized Bayesian Inference at the Bayes on the Beach conference in Australia next week.
Let's catch up if you're there!
@QUTDataScience
#BotB2024
I had a great time at
#AISTATS2023
with many interesting talks and conversations!
Big thanks to the organizers
@aistats_conf
@aistats2023
for putting together such a memorable event in oh-so-beautiful Valencia 🏝️
I'm teaching an intro workshop on Scientific Python this week 🐍
The target audience has a background in statistics, R, and cognitive modeling.
Which topics would you like to see covered in such a workshop? Any tips?
#rstats
#python
#AcademicTwitter
I'm currently preparing an article for the experimental track of
@JournalOVI
.
That means: The article is a Quarto document on GitHub, deployed via GitHub Pages. You can add collapsible paragraphs, callouts, Shiny widgets, videos, ... sky is the limit.
It's an absolute pleasure!
TIL that you can use purrr::partial() to pre-fill function arguments in R 🤩
That's one of my favorite python features and I'm genuinely happy to see it work in R!
The PyVBMC documentation is divine.
I'm rebuilding the entire documentation of the BayesFlow library for amortized Bayesian inference, and I'm constantly drawing inspiration from the great docs that
@AcerbiLuigi
@BobbyHuggins16
put together for PyVBMC👏
I crafted this in pure matplotlib a few weeks ago. It looks simple at first, but it made me go crazy. It’s a combination of:
- uncertainty bands or whiskers, depending on data stream
- broken x-axis
- zoom-in with larger markers and grid
🫠🫠🫠
@DavidKButlerUoA
By regrouping to a 5*N rectangle + x:
- blue fits snuck in blue
- orange fits snuck in orange
- result is a 5 * N rectangle (didn't count N)
- the red dot is left
-> total number is 5N + 1, which is not divisible by 5
(never had to count beyond 5, hope that's ok)
@tcarpenter216
I love statistical rethinking, but it might be too advanced when folks have no previous experience.
To get some exposure, I would not recommend textbooks, but rather online courses or videos. There are heaps of offers across those platforms like Coursera etc.
Great, that would successfully combine the worst of frequentist and Bayesian stats 😄 seriously: We combine the best of both worlds, and analyze sampling distributions of Posterior Model Probabilities (~ BFs on >2models) …but in a consistent Bayesian way:
🎥 My second YouTube video is live! 🔴
🤗 I talk about my personal experience with machine learning conferences as an early career researcher.
✨ I share 7 tips that I would have loved to know 3 years ago.
💭 What did I miss?
🍿 Watch the video here:
@rmkubinec
If you were to ask me again and again, the odds of me saying that I am a Bayesian (or something more extreme) even if I were a Frequentist are 5.3%.
Therefore, I was not able to reject my null hypothesis that I am a Frequentist at an α-level of 5% 🤷
I presented our work on data-efficient amortized Bayesian inference via self-consistency losses at the
@ELLISRobustML
workshop 🇫🇮☀️
Great crowd and interesting discussions!
Joint work with
@desirivanova
, Daniel Habermann, Ullrich Köthe,
@paulbuerkner
and
@StefanRadev13
.
@JimGrange
If this means that I never ever ever ever have to code a Drift Diffusion Model myself again, I’ll pay any amount of money for ChatGPT Plus 😅
@StefanRadev13
In 2020,
@paulbuerkner
was the first independent
#jrgl
to start in SimTech. 3 years and a research group of 5 doctoral researchers later, he is out and about to leave SimTech to become a Full Professor for Computational Statistics at
@TU_Dortmund
. Congrats and all the best! 🥳👏
Pushing the envelope of Bayesian surrogate modeling & simulation-based inference 🚀
Excited to finally release the pre-print of this new project with
@StefanRadev13
, Valentin Pratz,
@uPicchini
, Ullrich Koethe, and
@paulbuerkner
.
👉Check out BayesFlow at
New pre-print of our latest major BayesFlow upgrade:
“JANA: Jointly Amortized Neural Approximation” for simultaneous surrogate modeling & simulation-based inference. Comes with cool features for amortized Bayesian workflows and interpretable diagnostics:
Update:
Thanks to the great stats community on x dot com the everything app, my post found its way to the person who took that exact photo yesterday. I just mindlessly hit "save photo" on my phone and couldn't find it anymore today.
Credits to:
That hasn't been the Google logo since 2015...
I think
@aistats_conf
needs to update their banner files! 😂😂
(Pic stolen from a separate thread by
@shakir_za
)