🟢 𝗢𝗽𝗲𝗻 𝘁𝗼 𝗪𝗼𝗿𝗸 — 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 (𝗥𝗲𝗺𝗼𝘁𝗲 / 𝗔𝘂𝘀𝘁𝗶𝗻) → rexdouglass.com 𝗛𝗼𝘄 𝗱𝗼 𝘆𝗼𝘂 𝗳𝗶𝗻𝗱 𝗮 𝟭‑𝗶𝗻‑𝟭,𝟬𝟬𝟬,𝟬𝟬𝟬 𝗲𝘃𝗲𝗻𝘁 𝗶𝗻 𝟭𝟬,𝟬𝟬𝟬+ 𝗵𝗼𝘂𝗿𝘀 𝗼𝗳 𝘃𝗶𝗱𝗲𝗼? You 𝘀𝘁𝗿𝗮𝘁𝗶𝗳𝘆 𝘁𝗵𝗲 𝗵𝗮𝘆𝘀𝘁𝗮𝗰𝗸. #ComputerVision #ActiveLearning #DatasetCuration #MLOps #VideoAnalytics #Manufacturing #AppliedScience

🟢 𝗢𝗽𝗲𝗻 𝘁𝗼 𝗪𝗼𝗿𝗸 — 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 · Remote / Austin 🎞️ 𝗦𝗹𝗶𝗱𝗲𝘀 tinyurl.com/yxr7k5f8 🚧 𝗣𝗿𝗼𝗷𝗲𝗰𝘁 Completely unsupervised segmentation + labeling of parts, tools, and people in manufacturing. I demo it on a random “How It’s Made” fire-extinguisher video. 🧰 𝗦𝘁𝗮𝗰𝗸 DINO · SAM-2 · Gemini · YOLO

🟢 𝗢𝗽𝗲𝗻 𝘁𝗼 𝗪𝗼𝗿𝗸 — 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 · Remote / Austin 🎞️ 𝗦𝗹𝗶𝗱𝗲𝘀 tinyurl.com/yxr7k5f8 🚧 𝗣𝗿𝗼𝗷𝗲𝗰𝘁 Completely unsupervised segmentation + labeling of parts, tools, and people in manufacturing. I demo it on a random “How It’s Made” fire-extinguisher video. 🧰 𝗦𝘁𝗮𝗰𝗸 DINO · SAM-2 · Gemini · YOLO

Help! ─ 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 open to work ─ Remote/Austin → www.rexdouglass.com With new tools, I've been cooking: ▸ Machine-vision pipelines — RIOS Intelligent Machines ▸ Information-extraction pipelines — Microsoft ▸ Interactive SMS surveys — Pantheon Insights

Help! ─ 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 open to work ─ Remote/Austin → www.rexdouglass.com With new tools, I've been cooking: ▸ Machine-vision pipelines — RIOS Intelligent Machines ▸ Information-extraction pipelines — Microsoft ▸ Interactive SMS surveys — Pantheon Insights

Some ML Engineer at YouTube must handle just kids home for the holidays unsubscribing their parents from all the right wing channels.

90% of machine learning is data cleaning.

Who turned America's military against Americans? I'm tracking 313+ instances of support/opposition to the use of armed forces in domestic law enforcement in a new dataset: "Public Positions on Militarization of Domestic Law Enforcement in the U.S." docs.google.com/spreadsheets...

Public Positions on Militarization of Domestic Law Enforcement in the U.S. t.co/P5HmsHF1fA

Just the direct material costs for each additional year of war to Ukraine are about $111.7B or just over half their entire GDP burned per year 🚩$50B (25% of prewar GDP) to wartime defense spending over peace time levels 🚩$61.7B (30.8%) in direct destruction of things/sectors of the economy by Russia

This is absolutely nuts: 🚩30% of Ukraine's population is either internally displaced, a refugee outside of the country, or under Russian occupation. 🚩About 14% of the population's homes have been destroyed. Each additional year of war destroys about another 3%.

Each additional year of war in Ukraine implies an additional: 7,953 Civilians Directly Killed 12,822 Civilians Directly Injured

The bargaining model of war only has a few parameters. I'm often annoyed that the cost of fighting receives less serious measurement than military weapons/strategy. Then I remembered I pay $700 a month for LLMs, so here's the first in a series on Ukraine's costs from fighting.

Recently had a debate about two-way fixed effects and how it pretends to magically solve causal inference problems by just changing to a different DV than the one we actually care about. Here's that explained visually. Here's raw annual city temperatures:

hahahahahaha

We are delighted to announce Reticulate 1.41! You can now simply declare your dependencies using `py_require()`, and Reticulate will handle the rest. This seamless experience is powered by uv, an extremely fast #Python package manager written in Rust. Learn more: posit.co/blog/reticul... #RStats

Big Shout Out to Andrew Heiss at Georgia State: an API for grabbing the USAID data: foreignassistance-data.andrewheiss.com USAID greenbook: foreignassistance.andrewheiss.com/greenbook.html

There's been some confusion about whether the new DeepSeek-R1 open source model includes a political propaganda/censorship layer baked in. It does. Here's what it does in its own words and the experiments to reproduce.

Yes, I talk about this issue in my blog post: www.robertkubinec.com/post/fixed_e... Stata has similar behavior. It's a very big deal with two-way FEs.

Only 6% of papers using SOEP data share their reproduction workflows. We need #openscience now @janmarcus.de onlinelibrary.wiley.com/doi/10.1111/...

UpSet won the #ieeevis InfoVis 10-year test of time award 🎉. Thanks to @ngehlenborg.bsky.social @henstr.bsky.social Romain Vuillemot, and Hanspeter Pfister for being an awesome team to work with on this. I wrote up my reflections on what made UpSet successful: vdl.sci.utah.edu/blog/2024/10...

And make sure to check out our demo application: huggingface.co/spaces/ciass...

Finally out in @polanalysis.bsky.social (w/ @stefan-scholz.bsky.social, @zacharyst.bsky.social, @keremoglu.bsky.social and Bastian Goldlücke): "Improving Computer Vision Interpretability: Transparent Two-Level Classification for Complex Scenes" Available #OpenAccess at doi.org/10.1017/pan....

Ah, I found what I was looking for! From Imai et al.'s "Unpacking the Black Box of Causality" (2011). This is coming from the angle of two experiments in which you identify A -> B and B -> C and why you cannot infer the average causal mediation effect from that.

🧵 I'll talk up a recent paper I loved working on. We show why meaningful replications need to be based on deeper theoretical understanding & stronger empirical foundations. It's a theory paper like our other metascientific work. 🧪 #metasci #philsci #stats royalsocietypublishing.org/doi/full/10....

New post on the 100% Christmas Interval -- after almost 8 years, it's time that we finally talk about CIs. www.the100.ci/2024/12/05/w...

“We don’t value software, data, and methods in the same way we value papers, even though those resources empower millions of scientists” 💯 www.statnews.com/sponsor/2024...

Looking forward to reading this! "The Landscape of Causal Discovery Data: Grounding Causal Discovery in Real-World Applications" arxiv.org/abs/2412.01953 #stats #mlsky #blueskai 📈📉

On OOD: arxiv.org/abs/2211.16158 and I especially love this one: openreview.net/pdf?id=XCS_z... For more optimistic reads on uncertainty, most stuff by Ian Osband on uncertainty for (Deep) RL scholar.google.ca/citations?us...

arxiv.org/pdf/2409.02628 For problems with ensemble methods proceedings.mlr.press/v162/lotfi22... For issues with ML metrics (and, for drama, this reply iclr-blogposts.github.io/2024/blog/cl...) This lab has a collection of works www.vanderschaar-lab.com/uncertainty-... (pretty biased of course)

Our JSS article is out! And now I get to focus on {marginaleffects} 1.0.0. Stay tuned. www.jstatsoft.org/article/view...

Lab Experiment for Kids #3 Steel Wool and 9v Battery Ok seriously, I need more experiment suggestions.

Have you seen this? www.taylorfrancis.com/books/mono/1... (not proximal, but looks cool nonetheless)

Looks like this might be worth a careful read: www.sciencedirect.com/science/arti...

Preregistration is a floor, not a ceiling: Use it to specify initial analyses with the understanding that you can and should go further when you see unexpected patterns in the data. Gelman & Brown journals.sagepub.com/doi/full/10....

“…we systematically examined articles in four top psychology journals (N = 956) and found that 21% included open-ended data coded by humans. However, only about one-third of those articles reported sufficient details to replicate or evaluate the validity of the coding process”

Step 1: I show you a cheap-ish lab experiment for kids. Step 2: You give me ideas for what we should try next.

And another post on causal inference and social media and its effect on mental health. DAGs make a prominent appearance. As always, feedback welcome (especially if you think I am wrong). sjwild.github.io/blog/2024/11...

I am running my statistical rethinking course again starting in January. But this time just for Leipzig locals, so I can work with a smaller group this year and track individual progress better. I think that will help me tune the material, espeically homework, better. Materials still open however.

Amazing demo by Vaibhav Srivastav of structured data extraction running on an LLM that executes entirely in the browser (Chrome only for the moment since it uses WebGPU). https://simonwillison.net/2024/Nov/29/structured-generation-smollm2-webgpu/

Pro Tip: Uninformative priors don't look anything like you expect in the sample space of the actual outcome. Don't just pick priors in the blind, simulate the full null model with just priors, and verify it makes predictions like you expected in your head.

🚨 New Paper 🚨 Can LLMs perform latent multi-hop reasoning without exploiting shortcuts? We find the answer is yes – they can recall and compose facts not seen together in training or guessing the answer, but success greatly depends on the type of the bridge entity (80% for country, 6% for year)! 1/N

I have a new piece out with @aisvarya17.bsky.social in @columjournreview.bsky.social in which we test how OpenAI's new search feature surfaces and attributes news content. Our findings were not promising for news publishers (1/9) www.cjr.org/tow_center/h...

2/ To kick off the thread, here are 22 persuasive communication devices that you should watch out for when writing or reading a research article. https://buff.ly/3Z9Mb5C

How does Approximate Bayesian Computation *ABC) work? @rabaath.bsky.social has a classic nice simple intro the ABC using the problem of infering the number of socks in a sample of laundry. There's a movie and a page with all the code. www.sumsar.net/blog/2015/07...

Ooooooo www.sciencedirect.com/science/arti...

Releasing: a dataset of two million Bluesky posts. This dataset has been collected using Bluesky's API, and I hope it will be useful for all the researchers out there!

The thing is, there's already a dataset of 235 MILLION posts from 4 MILLION users available for months. Not sure why @hf.co is a target of abuse zenodo.org/records/1108...

Maybe this will help? arxiv.org/abs/2401.05330