
MSc @Skoltech. Open Source fan at @pymc_devs, Geoopt. @pymc_labs ninja. Also a great fan of Hyperbolic DL. Love rollerscating, skiing.
Moscow, Russia
Joined December 2016
- Tweets465
- Following194
- Followers853
- Likes1.2K
Pinned Tweet
📊Conformal prediction or Bayesian inference to measure uncertainty? The answer depends on which problem you are dealing with:
- Is it black-box?
- Is it white-box?
To get the difference, read my latest blog post:
ferres.ml/en/posts/2023/prob…
This is massive
Any PyTorch users out there that are interested in contributing to #PyMC? I asked GPT to port our PyTensor JAX-backend to PyTorch and this is what it came up with: github.com/pymc-devs/pytenso… It doesn't work yet but it's a good starting point and there are (failing) tests.
3
594
A crazy idea, there should be something trained on Wikipedia to create auto tags!
One of my subscribers commented on my article on the Gzip + KNN paper [open.techwriters.info/codeco…]
They created their own implementation of the classifier using zstd. His implementation is blazing fast and achieves high accuracy.
He employed a neat trick. When using zstd, you can build a dictionary out of your data which can later be used to compress any other data.
He created dictionary for each class in the training dataset and then built a compressor using that dictionary. Essentially he built one compressor per class.
In order to classify a new piece of text, he would compress that text using each of the class specific compressors and pick the class of the compressor which achieved maximum compression.
The basic idea is that, a document from a given class will be maximally compressible using a dictionary built from the documents of its actual class.
This might not work as well on out of distribution data, because the dictionary built is specific to the training data and out of distribution data may have different words and patterns not seen by the pre-built dictionary. But still very impressive and innovative.
Here's a link to his code: github.com/cyrilou242/ftcc
1
315
Brilliant
Simple and powerful math ideas 💡 will always beat non math ideas.
In the new paper a simple model based on gzip and KNN obliterated transformers with hundreeds of million parameters.
How did it do it? The central powerful idea is that if x1 and x2 are similar their joint compression is more efficient than compressing them separately.
#llms #transformers #timeseries
330
Maxim Kochurov retweeted
Looking for a deeper understanding of GPs and their practical applications in time series analysis❓
Don't miss the closing webinar session where @ferrine96 will explore essential concepts and demonstrate the Rolling regression example from pymc-examples.
meetup.com/pymc-labs-online-…
Maxim Kochurov retweeted
1/ Are you ready to unlock the power of Gaussian Processes? Join us for an engaging webinar where we dive into the fascinating world of Gaussian Processes and their applications.
📅 June 22nd, 2023
⏰ 9am PT / 12pm ET
🎙️ Speaker: @ferrine96
🔗 Register:👉meetup.com/pymc-labs-online-…
Maxim Kochurov retweeted
This is actually how "confidence intervals" work in statistics. I'm being 100% serious. Read on. This is a really fun concept.
nitter.cf/HumansNoContext/status…
🔥Linear Regressions can be OP🔥
With Bayes you can inform:
📊One feature is more important than another
🏹The extent to which linear regression is accurate
➕The sign of the effect size
Today, State of Bayes (May 25, 16:00UTC+0):
meetup.com/pymc-labs-online-…
Do not miss the 3d session of the State of Bayes webinar. I will talk about hierarchies, where they come from and how to move from a non-hierarchical model to a hierarchical one. We will also learn how to effectively use prior predictive to debug a model
meetup.com/pymc-labs-online-…
With the teaser left behind, don't miss the first full lecture in the webinar series "State of Bayes," tomorrow, April 20! You will learn the core principles behind the Bayes formula and how it all relates to real life
Use this link to register:
meetup.com/pymc-labs-online-…
Starting next month I'll read a free webinar series about practical Bayesian methods we use in @pymc_labs. Schedule your time and invite anyone you know! meetup.com/pymc-labs-online-… #PyMC
I was struggling to formulate when Bayesian ab testing supersedes even to myself. @PhDemetri wrote an excellent thread about the real motivation of being and not being Bayesian.
Lots of discussion about Bayesian AB testing being better than frequentist.
Look, I don’t disagree. Bayes is my preferred method, but there are a few things I think we need to get clear on. 🧵
3
449
Maxim Kochurov retweeted
The AB Testing // Causal Inference // Econometrics // Statistics cross over language kills me
Curious about heteroscedasticity in your data? The Bayesian way is made easy by the new Bambi release
I'm so happy to announce Bambi 0.10.0 was released.
This release comes with an exciting new feature: Bambi now supports GLMMs for location, shape, and scale.
See the example: bambinos.github.io/bambi/not…
If you want to know more, keep reading!
8
779
Replying to @tomicapretto
@tomicapretto recommended me a paper to read and it is worth sharing. For linear regression it allows you to express priors in the form
* Few variables contribute to prediction
* Feature A is more important than B
* My regression should have low R^2
ar5iv.labs.arxiv.org/html/22…
The blog post and the notebook for the tweet
ferrine.github.io/posts/2023…
github.com/ferrine/ferrine.g…
Last Autumn I’ve held my first course “Practical Bayes” at Moscow State University. You can find there hierarchical models, AB tests, GPs and CLV. Here is the page with all the slides at one place ferrine.github.io/pages/prac…
Maxim Kochurov retweeted
PyMC supports JAX for fast CPU and GPU sampling. But how can you use JAX functions directly in your @pymc_devs model?
In this post, @RicardoV944 and @aseyboldt show how you can integrate JAX to build fast Bayesian Neural Networks and ODEs.
pymc-labs.io/blog-posts/jax-…
Maxim Kochurov retweeted
We are adding a new cool decorator in the Memray pytest plugin that will make your test fail if it leaks memory 🤖 Is like valgrind but a LOT of times faster 🚀and it doesn't get confused by the Python interpreter specialised allocator. And configure it is just one line! 🙀


















