Image models struggle to create Go boards correctly. I thought ChatGPT may be able to copy a correct one into a scene, but conditioning the model is through too small a bottleneck. "Please copy this image exactly to create a new image. Every line and stone should appear in exactly the same places."

We were fortunate to be joined by two new machine learning faculty in Edinburgh Informatics this year. Both are looking for PhD students:

Viacheslav Borovitskiy (Slava), geometric learning or uncertainty quantification: https://vab.im/vacancies/

Nikolay Malkin, Bayesian machine learning (including probabilistic reasoning in language, generative models, and neurosymbolic methods) and applications in the sciences: https://malkin1729.github.io/

Some funding available to support probabilistic AI activity in the UK. For events, public engagement, visits, and 1-year bridging postdocs: https://probai.uk/funding-1 -- more in the Prob_AI hub newsletter: https://www.probai.uk/news

@akkana glad it's useful. I've really appreciated several things on your blog over the years (including an old firefox middle click paste fix, mutt instructions, gimp saver-as).

The University of Edinburgh is looking for someone to direct a new generative AI lab. It's a professor role, with lots of resources. More info: https://www.aspenpeople.co.uk/uoe-gail/

For earlier-career faculty jobs (and a communications role) see: https://www.ed.ac.uk/informatics/about/work-with-us/vacancies

Machine Learning faculty job hunters: there's an open position in Edinburgh. Hard deadline 15 Jan 2024 (UK time).

Full details: https://elxw.fa.em3.oraclecloud.com/hcmUI/CandidateExperience/en/sites/CX_1001/job/8801

A thread by a colleague on why Edinburgh is awesome: https://nitter.net/oisinmacaodha/status/1720142344618451193
I've stolen the photo he used of the inside of our building.

@pronoiac Apologies for the slow reply. That looks fine to me (after removing the `;` to make it work). A previous version of my bookmarklet could be tricked into setting the `document.location` to something starting with `javascript:`, which can't happen with this code.

Mastodon has content security set such that the javascript wouldn't run anyway. But potentially if someone tricked you to trigger my old mastodon bookmarklet on a non-mastodon page there could have been a problem.

All major machine #translation systems #translate the Chinese sentence 她每天都不吃早饭。incorrectly to “She doesn’t eat breakfast every day”. Does #ChatGPT get this difficult case wrong too? Yes.

If asked to be careful, or to reason about the sentence, does ChatGPT give any indication that there's a problem? Usually not!

However, discussing the sentence in Chinese does work. And with the right prompting, it is possible to notice and then fix the issue.

Full exploration here: https://homepages.inf.ed.ac.uk/imurray2/pub/22chatgpt.html

And now fixed the bookmarklet so it can toggle to/from an instance located somewhere other than the domain in the username.

But the bookmarklet's still pretty hacky/fragile.

@sethaxen I agree, a monotonicity constraint isn't compatible with being Gaussian. But yes, you can strongly encourage a GP to be monotonic: http://proceedings.mlr.press/v9/riihimaki10a.html

If you just want a flexible model over monotonic functions, you could use a positive combination of monotonic basis functions. Or a spline constrained to be positive, as we used as part of this work: https://arxiv.org/abs/1906.04032 with code for the representation in several libraries, including https://github.com/bayesiains/nflows

@nbranchini While not quite what you were getting at, slide 18 of my 2009 MLSS tutorial mentioned two well-known forms of exponential dimensionality dependence (inspired by bits of MacKay's book):

https://homepages.inf.ed.ac.uk/imurray2/teaching/09mlss/slides.pdf#page=18

@nbranchini Nice explicit example, and thanks for the citation 🙂 .

The overarching message of my chapter was meant to be: Monte Carlo can be really good. For positive functions, we could even form Monte Carlo estimates of integrals with zero variance (bottom of p23). However, in high-dimensions it's hard to sample from an appropriate distribution, even if you've found one. (And finding a suitable distribution is also hard in high dimensions.)

I should have connected up the dots better.

@avehtari I knew about BDA, but didn't know "Regression and Other Stories" was also free online. That's amazing; thanks!

All three books I've co-authored are freely available online for non-commercial use:

#Bayesian Data Analysis, 3rd ed (aka BDA3) at https://stat.columbia.edu/~gelman/book/ and lectures plus #rstats, #Python and #Stan code at https://avehtari.github.io/BDA_course_Aalto/

#Regression and Other Stories at https://avehtari.github.io/ROS-Examples/ including #rstats and #Stan code

Active Statistics at https://avehtari.github.io/ActiveStatistics/

The websites also have links to the publishers' web stores if you prefer hard copies of these

It's unfortunate that a different version of the bookmarklet is needed depending on whether you are on a recently-upgraded Mastodon instance or not.

To avoid this split in future, I filed an issue about the problematic (broken) redirect behavior in v4:
https://github.com/mastodon/mastodon/issues/20145
[update: and a fix was committed in less than a day! It will take a while to filter through to all live v4 instances, a downside of distribution.]

@MilesCranmer Hopefully it more-or-less works for you. There's lots of annoying edge cases between Mastodon versions, which I didn't *entirely* sort out. I'll be interested to see some reports either way.