Kyle Cranmer
@kylecranmer.bsky.social
Director Data Science Institute @UWMadison, Professor of Physics,
EiC @MLSTjournal. Physics, stats/ML/AI, open science.
EiC @MLSTjournal. Physics, stats/ML/AI, open science.
It is Inherently vague and ambiguous
November 10, 2025 at 7:49 PM
It is Inherently vague and ambiguous
Hopefully, my flights don't get delayed or canceled.
November 9, 2025 at 8:43 PM
Hopefully, my flights don't get delayed or canceled.
I got tired of mashing together tools to write long threads with 𝐫𝐢𝐜𝐡 𝑓𝑜𝑟𝑚𝑎𝑡𝑡𝑖𝑛𝑔 and ℳα†ℏ—so I wrote La𝑇𝑤𝑒𝑒𝑡!
It converts Markdown and LaTeX to Unicode that can be used in “tweets”, and automatically splits long threads. Try it out!
keenancrane.github.io/LaTweet/
It converts Markdown and LaTeX to Unicode that can be used in “tweets”, and automatically splits long threads. Try it out!
keenancrane.github.io/LaTweet/
November 3, 2025 at 6:35 PM
We made an automation that runs on MacOS (and there is a Linux version) that does system wide latex -> Unicode through a hot key
www.unicodeit.net
www.unicodeit.net
unicodeit.net
www.unicodeit.net
November 3, 2025 at 6:32 PM
We made an automation that runs on MacOS (and there is a Linux version) that does system wide latex -> Unicode through a hot key
www.unicodeit.net
www.unicodeit.net
It's been a very fun project and I've really enjoyed working with my collaborators @rkhashmani.me @marielpettee.bsky.social @garrettmerz.bsky.social and Helen Qu (with one “l”, sorry about the typo in first tweet)
October 28, 2025 at 5:28 PM
It's been a very fun project and I've really enjoyed working with my collaborators @rkhashmani.me @marielpettee.bsky.social @garrettmerz.bsky.social and Helen Qu (with one “l”, sorry about the typo in first tweet)
We also add a target variable θ∈ℝ that has known mutual information with the data sets x₁, ..., xₙ so that one can test how well a down-stream model with a shared representation can predict θ.
This is all aimed at making benchmark datasets and for powering ablation studies
This is all aimed at making benchmark datasets and for powering ablation studies
October 28, 2025 at 5:23 PM
We also add a target variable θ∈ℝ that has known mutual information with the data sets x₁, ..., xₙ so that one can test how well a down-stream model with a shared representation can predict θ.
This is all aimed at making benchmark datasets and for powering ablation studies
This is all aimed at making benchmark datasets and for powering ablation studies
We also use simple linear structural equations to tell a causal story with hyper-parameters that are easy to interpret. So we start with some "proto-latents" u that are totally uncorrelated and then construct the latents z through linear combinations. Here ũ₁ is a common cause
October 28, 2025 at 5:23 PM
We also use simple linear structural equations to tell a causal story with hyper-parameters that are easy to interpret. So we start with some "proto-latents" u that are totally uncorrelated and then construct the latents z through linear combinations. Here ũ₁ is a common cause
Each flow can be pre-trained to generate realistic looking data for a given setting... e.g. images, scientific data, etc. We used images for this demonstration.
We know the mutual information between this set of frogs and cars!
We know the mutual information between this set of frogs and cars!
October 28, 2025 at 5:23 PM
Each flow can be pre-trained to generate realistic looking data for a given setting... e.g. images, scientific data, etc. We used images for this demonstration.
We know the mutual information between this set of frogs and cars!
We know the mutual information between this set of frogs and cars!