astanziola

We trained a foundation model on 18 million heart ultrasound videos to predict structure instead of pixels. Introducing EchoJEPA, the first foundation-scale JEPA for medical video. Paper: arxiv.org/abs/2602.02603 Code: github.com/bowang-lab/Ech… 🧵 1/n

New manuscript shows how to extend automated model discovery and universal differential equations to chaotic systems in #neuroscience using a trick from control literature known as the Prediction Error Method (PEM)! Paper: arxiv.org/abs/2507.03631 #sciml #ai4science

Making good on this promise—in the fastest turnaround time ever—my collaborator Etienne Corman has just posted MATLAB code for #RectangularSurfaceParameterization here: github.com/etcorman/Recta… (C++ version is still in the works…)

@value_and_grad @giladturok @giffmana I agree that the docs are not at the level of pytorch (especially because the community is not that big), but what's missing from jax that you think is kept internal?

@giladturok @giffmana It's kind of Google's tradition to churn out the best tooling ever with no docs, while keeping the internal version separate from the open-source one. It almost feels like they're just giving you a taste of what they really use to lure you in. tbh I’d join Google just to use JAX.

So this was published in nov 2024, and looked like pytorch is sloooowly getting to jax levels of niceness. Now half year later, o3 said it's in nightly, to be released in 2.10, but that's a big fat hallucination. Is the plan to properly get it into pytorch at some point?

@docmilanfar Disagree, this one is gold: amazon.it/Teoria-dei-seg…

nearly all undergraduate signal processing textbooks

💫Fellow scientific computing geeks, here is an example of adaptive mesh refinement in #JAX: gist.github.com/patrick-kidger… I recently heard a claim that this was impossible in JAX due to its lack of dynamic shapes, and decided to prove the naysayers wrong 😄

There has been significant recent interest in methods that use random walks to solve PDEs. In a project to be presented at SIGGRAPH Asia (w/Ekrem Yılmazer and @DelioVicini), we investigated how to solve *inverse PDE* problems by differentiating such solvers.

How can we train biophysical neuron models on data or tasks? We built Jaxley, a differentiable, GPU-based biophysics simulator, which makes this possible even when models have thousands of parameters! Led by @deismic_, collab with @CellTypist @ppjgoncalves biorxiv.org/content/10.110…

With this in mind, the spectral regression that Sander is talking about isn’t a linear regression in the standard Fourier domain. It is a regression onto a basis of eigenvectors of the weight matrix W that are *adapted* to the content of the image and the training data. This is more like practical harmonic analysis than your typical spectral autoregression model. For reference, these ideas are described in these two papers. 4/4

Our department is currently advertising for two new Lecturers (Asst Professor with tenure). These positions don't open that often, so please circulate widely! The posts are open to all of Medical Physics and Biomedical Engineering (including ultrasound!) jobs.ac.uk/job/DJI940

Left: Reinforced diffusion transformer splatting with deep equivariant flow matching Right: Cubic splines

📢 ⚡️Glad to finally release #NAVIX: we re-implemented MiniGrid in #JAX for lightning-fast RL training🚀 You can now train more than 2048 #PPO agents on #MiniGrid environments in less than *2* minutes! ⚡️⚡️ Say goodbye to long and tedious hparams search! 👋👋

📢 Come chat about #creditassignment with #languagemodels tomorrow at #ICML2024 ! 📅 27 Jul, 10am 📍Room Stolz 0

We enjoyed three excellent presentations on #focusedultrasound simulations and new systems in the first session of #FUN24. Thanks to Elly Martin @ellymmartin, Will Grissom @wgrissom, Bradley Treeby for their insights! @cfcaskeyPhD @huiwen_luo @ian_simcock @TulikaNan @IoanaGrigoras4 @cjstagg @UCL_Ultrasound

The talk, slides, and code for our SGP graduate school course on MCGP can be found here: rohan-sawhney.github.io/mcgp-resources/ We cover a lot of ground: Monte Carlo integration, differential equations, Brownian motion, and of course walk on spheres, which nicely ties all these topics together!

📢! Tomorrow (27th) I will be presenting our - *Survey on Credit Assignment* and our work on - *Credit Assignment with Language models* (SPOILER🚩) at @UoE_Agents (@InfAtEd) reading group. Thank you, @UoE_Agents for the kind invite! Tune in, if interested! (links below)

All unit tests passed.

@HarveyHCodyII @ClimateBen Suboptimal according to what metric?

"Scientists now know today's capitalist economic system can't limit global warming to well below 2°C." That is very likely true. Also true is that a capitalist system is the best economic system humans have devised. So, if your claim about our system is true, the system that must be created will be suboptimal. Couple that with the fact that much of what people now have or use will be banned; the likelihood that any country can afford to build what is needed for Net Zero is near zero. What you propose appears to be insane.

Scientists now know today's capitalist economic system can't limit global warming to well below 2°C meaning human adaptation will not be feasible as the conditions for modern agriculture disappear forever in the coming years. Most scientists accept this. Some aren't ready yet. 🧵

@MichaelAArouet @petelgeuserc Italy from post WW2 until the change towarsa a free-market led economy and the privatization of its main assets and institutions, started in the beginning of the 90s.

@petelgeuserc Show me one example of socialist country not ending up in poverty, misery and terror against own citizens

@katsen @kaseyklimes The last thing the world need is real life SEO

Put it in reverse... Imagine the economic arbitrage opportunity for local entrepreneurs and businesses that figure out how to get lower-rent streets to be marked 'scenic' by Maps. Another hypothesis - the downside may actually be quite low. Do 'speed'-based drivers contribute economically when just passing through? Do 'scenic' drivers often stop if they're getting poor recommended routes (eg - not scenic)? Relevance systems become powerful marketplaces if you let them do their jobs. Nearly every Google product has proven this out again and again, no?

this got talked about a lot when I worked on google maps. I would personally love this feature, and it’s technically feasible. here’s why I fought it loudly every time it came up 🧵