Jan Lause 🟦 @janlause.bsky.social🦉

data science PhD student 🧬🧠🖥️ scRNA data, retina & brain | sci-comm 🤖❤️🌍 AI & sustainability | easily sidetracked by small details in plots & birds 📈🔍🦜

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

In their Specious Art paper, Chari & @lpachter claim that tSNE/UMAP are as arbitrary as a random elephant shape. But are they?

We show in our comment that this is false and throws the tSNE/UMAP baby out with the bathwater!

Details in 🧵& paper:

https://t.co/EyINxIcurb

1/8

[Lior Pachter @lpachter]

It's time to stop making t-SNE & UMAP plots. In a new preprint w/ Tara Chari we show that while they display some correlation with the underlying high-dimension data, they don't preserve local or global structure & are misleading. They're also arbitrary.🧵https://t.co/XkAOTKlOcs

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

PS: If anyone needs additional reasons to leave this place, here is one: Musk endorsed the German right-nationalist AfD — a party that spreads hate against migrants and constantly lies about climate change.

Unacceptable!

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

I won't be reading, posting or replying here in the foreseeable future. Find me (and many other science & sci-comm people) on the other site, where skies are still blue! 🦋

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

I won't be reading, posting or replying here in the foreseeable future. Find me (and many other science & sci-comm people) on the other site, where skies are still blue! 🦋

[Jeremy Howard @jeremyphoward]

Narrative on X: 🦋 has no AI/ML and just talks about itself
My actual feed on 🦋:

[Gary Marcus @GaryMarcus]

This other site is starting to gel. You can find me there.

[bryan dickinson @chembioBryan]

I am very skeptical of the AI/ML/computational methods to predict protein-protein interactions. Prove me wrong.

I present a challenge for anyone who claims they can predict protein-protein interactions.

1/n

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

@Ricarda_Lang @Die_Gruenen 👋

[Sandra Wachter [email protected] @SandraWachter5]

Another example of what I @b_mittelstadt @c_russl termed careless speech. Subtle hallucinations are dangerous & developers are not (yet) liable for them. We argue they should. See paper: Do LLMs have a legal duty to tell the truth? https://t.co/oVM3etcaJZ https://t.co/GMksxb5FqE

[Dmitry Kobak @hippopedoid]

We updated our ICLR dataset (see https://t.co/Loo2f149aQ) with blind 2025 submissions to @iclr_conf. Over 10k submissions this year.

I like how this embedding (SBERT + tSNE) shows which ML areas are old-school and which ones are currently booming.

[Anthropic @AnthropicAI]

Over the past few months, our Interpretability team has put out a number of smaller research updates. Here’s a thread of some of the things we've been up to:

[Yin lab @HenryYin19]

The vocabulary of systems neuroscience may appear daunting to many. Here's a short dictionary of common terms. BTW if you use them in your papers and grants you will have greater success

[Ricarda Lang @Ricarda_Lang]

Warum lügt die @CSU?

Kurz gesagt: Um von echten Problemen abzulenken, also um euch zu verarschen.

Also kuschelt eure Haustiere, widersprecht so einem Mist und lasst uns die Probleme lösen, die euch im Alltag stören und nicht welche, die die Union erfunden hat 🐶😽❤️

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

@KordingLab We are! 👋 While we acknowledge limitations (distance preservation is bad, not all high.dim. neighbors stay close, misuse is possible), tSNE/UMAP do preserve many neighbors&clusters which is useful for exploration/sanity checks.

Full write up ⬇️

https://t.co/zOm1IPOZtI

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

In their Specious Art paper, Chari & @lpachter claim that tSNE/UMAP are as arbitrary as a random elephant shape. But are they?

We show in our comment that this is false and throws the tSNE/UMAP baby out with the bathwater!

Details in 🧵& paper:

https://t.co/EyINxIcurb

1/8

[Kording Lab 🦖 @KordingLab]

Sequencing folks. You seem to use UMAP all the time. But everyone is making fun of it. Someone must be willing to defend it?

[Tom McCoy @RTomMcCoy]

🤖🧠NOW OUT IN PNAS🧠🤖

Language models show many surprising behaviors. E.g., they can count 30 items more easily than 29

In Embers of Autoregression, we explain such effects by analyzing what LMs are trained to do

https://t.co/lJIWx89YpJ
Major updates since the preprint!

1/n

[Leland McInnes @leland_mcinnes]

Datamapplot 0.4 is out now, and has far more powerful and effective interactive plots.
Here is an example of a Data Map of 2.4 million papers on ArXiv, ready to be explored.

[Dvir Aran @dvir_a]

Nobel prize for t-SNE!

[NimwegenLab @NimwegenLab]

The strange state of current publishing. I think some stats on our recent Cellstates paper provide interesting food for thought. The preprint was put on BioRxiv and within 3 months there were ~8500 abstract views and >2000 PDF downloads.
We also submitted it to PCB 6-11-2023 1/n

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

@MicTott I agree that t-SNE/UMAP are (only) visualization tools that distort the original space, but they are not "arbitrary", even though some people have claimed that.

We recently wrote a comment on why exactly that claim is too strong:

https://t.co/zOm1IPOZtI

[Jan Lause 🟦 @janlause.bsky.social🦉 @JanLause]

In their Specious Art paper, Chari & @lpachter claim that tSNE/UMAP are as arbitrary as a random elephant shape. But are they?

We show in our comment that this is false and throws the tSNE/UMAP baby out with the bathwater!

Details in 🧵& paper:

https://t.co/EyINxIcurb

1/8

[Melanie Mitchell @MelMitchell1]

PSA:

If you don't want your X posts used to train Grok, you now have to explicitly opt-out.

Go to https://t.co/V8hWTOOXAs and uncheck the box.

If link doesn't work, go to Settings->Privacy and Safety->Grok

[Dmitry Kobak @hippopedoid]

ICML 2024 (held in Vienna) registrations vs. registrations *per million inhabitants* by country.

The barplot of registrations on the log scale was shown yesterday during the opening. I took a photo, digitized with WebPlotDigitizer, and normalized per capita.

#ICML2024 @icmlconf