It feels like a tipping point has been crossed where the butterfly site has more interesting content than the everything app. You can find me there under the username adrhill!.

I have some FOMO about not tweeting our latest preprint here, but you know where to find it.

Working on this talk inspired us to develop .a unified interface (and testing/benchmarking system) for over a dozen Julia AD backends, which deserves its own thread!.

The goal of our talk is to give both of these groups the right tools to make informed decisions on how to use AD in Julia.

Additionally, we identify two different types of AD users:.1. package developers want to make their functions differentiable with as many backends as possible.2. package users want easy ways to compute derivatives with the best possible performance.

Should you use generic types? Can you perform mutating operations? Newcomers may not understand where these constraints come from, or whether they are inevitable.

Meanwhile on the Julia side, the dream is to make the whole language differentiable. Ideally, you should be able to write the code you want, and get gradients for free. In practice, there are multiple AD packages with distinct tradeoffs.

In Python, AD is fragmented into separate frameworks like TensorFlow, PyTorch or JAX. Each one of these has its own package ecosystem, and interoperability is limited. You first pick your AD framework, and then you write all your code in it.

Are you curious what automatic differentiation looks like in Julia? The talk @giomdal and I gave at @JuliaConOrg is finally online: 🧵 1/7.

Accidentally used his GitHub username: the correct Dalle is of course @giomdal. 😅.

Also fun to spot DifferentiationInterface and DitherPunk! :)

I never thought work of mine would be shown in a "State of Julia" talk, let alone in a football stadium, but with @gdalle and @jmmsmit made it happen! #JuliaCon2024

RT @ml_tuberlin: We have a new website 🎉 . Thanks to all the people who put countless hours into making it look as amazing as it does! . W….

Many thanks to Fons @PlutoJL, @theo_gf, @niklasschmitz_, @ml_tuberlin who helped make this happen!.

All of it is MIT licensed. Some additional behind-the-scenes information and discussion can be found in the announcement post on the Julia discourse:.

Lectures and homework are written using @PlutoJL notebooks, giving students immediate reactive feedback while working on exercises.

I highly recommend the lecture on Automatic Differentiation that covers forward- and reverse mode AD and packages in Julia's diverse ecosystem.

The course doesn't assume any prior knowledge of Julia. Fundamentals are covered in the first three lectures, followed up by lectures on Automatic Differentiation and Deep Learning.

I already accidentally announced it on Friday, but it deserves its own tweet: . This is a 5 week course on Julia's excellent Machine Learning ecosystem. The goal: teaching you to to write reproducible, unit-tested Julia code and do Machine Learning research in Julia.

I'm proud to announce the "@JuliaLanguage for Machine Learning" course at @ml_tuberlin!. GitHub: Website: A thread 🧵 1/7.