Steven Atkinson

Statistics, machine learning, physics, engineering, scientific computing, music, woodworking.

[Steven Atkinson @SASExperience]

Hey, Twitter! New work on discovering physics from data--with uncertainty quantification! Answering the question "What do our data tell us about physics--and what *don't* they tell us?" Bayesian Hidden Physics Models: https://t.co/HCRGJh1UML

[Steven Atkinson @SASExperience]

@aidangomezzz Psh, everyone knows that 3e-4 is the best learning rate for Adam, hands down.

[Steven Atkinson @SASExperience]

@mark_riedl Do you think it's relevant that (1) we don't know *what* ChatGPT is (lack of precise info from OpenAI) and (2) we can't guarantee access to it (what if they pull the plug tomorrow?)

Feels like putting an anonymous pen pal as a coauthor. Related to accountability?...

[Steven Atkinson @SASExperience]

@savvyRL PINNs would fit!

[Steven Atkinson @SASExperience]

@vboykis It was also helpful because I got a (reasonably correct!) description by example of the WAV file structure that made subsequent deep diving way easier.

So it was very helpful where I had gaps. Probably less helpful where I'm more experienced.

[Steven Atkinson @SASExperience]

@vboykis I used ChatGPT to help me write some C++ to read in a WAV file.

* Seems boilerplate
* Googling was surprisingly unhelpful
* Its code was more idiomatic than mine (I haven't really written C++ in a long time).

It made some errors that I fixed as I ported its answer.

[Steven Atkinson @SASExperience]

"licensing 101" should be part of technical education.

In grad school, I put my code on GitHub (with no license) and thought that meant that I had open-sourced it.

[Jo Kristian Bergum @jobergum]

“Open sourcing everything”

This is misleading, CC BY-NC 4.0, which Meta often uses, is not an open source license.

[AK @_akhaliq]

Riffusion, real-time music generation with stable diffusion

@huggingface model: https://t.co/on5deEcQS1
project page: https://t.co/FcPhFpzKmB

[Steven Atkinson @SASExperience]

I suspect that the distribution of chats people are having with ChatGPT is quite different when they're not making content to share on the internet.

[Steven Atkinson @SASExperience]

I enjoyed watching this. Fantastic pedagogical style on a profound class of problems combining stat phys with computation. I admire how well Lenka has bridged that gap and argues its significance.

[Lenka Zdeborova @zdeborova]

Happy to share the recording of my @EPFL_en Inaugural lecture titled "The Physics of Algorithms". Discover phase transition in computational problems and their relation to algorithmic hardness, in 30 min, rather general public style. https://t.co/PdfAp0iIsP

[Steven Atkinson @SASExperience]

Feels like the noise level is going up in Twitter's notifications (eg got a push notification for cold spam).

Thinking I'll turn them off.

[Steven Atkinson @SASExperience]

@RhysGoodall Do you mean that samples x have to lie on a simplex? For cases like that I usually pick points in $z \in \R^d$ s.t. x=softmax(z); the metamodel can still be fit as f(x).

Can't say I've seen any papers about it, though. Is there an aspect you're looking for that this is missing?

[Steven Atkinson @SASExperience]

@AlphaSignalAI @omarsar0 @ylecun Black-box generative modeling is not a feasible solution for a scientific writing tool.

Galactica is like if Google Maps populated their maps with a generative model. You wouldn't say "Oh well, it wasn't perfect" if it were wrong; you'd say "Why use a generative model for this?"

[Steven Atkinson @SASExperience]

@addisonsnell My local paper does reprints of the NYTCW a few weeks late, so I was excited to see this come up this week. Looking forward to working on it!

[Steven Atkinson @SASExperience]

@unsorsodicorda @Branchini_Nic (Full disclosure, didn't check closely to see if the Gaussian was full rank or had a nice decaying spectrum...but assuming the former?)

[Steven Atkinson @SASExperience]

@unsorsodicorda @Branchini_Nic Right... But do I care about Gaussians? Lots of "high-dimensional" functions have eg active subspaces on a domain of interest. Or eg PCA -> KLE remains pretty tame for "classic" UQ problems, and the physical insight holds up in practice for eg many heterogeneous materials?

[Steven Atkinson @SASExperience]

@ducha_aiki Sure. I personally find myself doing lots of problems where the train set doesn't include y's that I want to be able to predict at test time, but I totally get how this can work in some domains (thinking big data/high-dim where every y_i will be seen plenty).

[Steven Atkinson @SASExperience]

@ducha_aiki > about what it's going to need to predict that a regression model wouldn't have. (Sort of doubt this part helps nearly as much, but it always weirds me out building a model that a priori couldn't predict some values... So maybe that's the trade off?)

[Steven Atkinson @SASExperience]

@ducha_aiki (Took me a bit to find your explanation: https://t.co/IVGikh24bJ)

My guesses: 1/2 only part of the model needs to be correct when predicting (bins that are wrong the model can output whatever--easy to learn) and (2/2) by discretizing you give the model additional prior info >

[Dmytro Mishkin 🇺🇦 @ducha_aiki]

@ChrisBridges95 @JFPuget @tunguz Quantize the possible range train bin clf. Output the center of the bin, or maybe do local softargmax

[Steven Atkinson @SASExperience]

Made @[email protected] 🙂