Eric Volkmann

13 posts

Eric Volkmann

Eric Volkmann

@e_volkmann

Linz, Austria Katılım Kasım 2024
35 Takip Edilen83 Takipçiler
Eric Volkmann
Eric Volkmann@e_volkmann·
@LL_AIlo @fchollet We did add custom kernels via the FFI interface of JAX, which gives us ~2x speed-up on our GPU. More is probably possible with more time investment. The annoying thing about custom kernels is that the performance typically doesn't transfer across different GPU generations.
English
0
0
2
15
Louis | AI Bullshit
Louis | AI Bullshit@LL_AIlo·
@fchollet Totally, JAX is a game-changer for performance. But from my experience, the real magic happens when you pair it with custom ops in TensorFlow. Those little tweaks can unlock optimizations that pure JAX might mi
English
1
0
3
90
Eric Volkmann
Eric Volkmann@e_volkmann·
@ggalletti_ Our initial goal was ML pipeline integration, but the process convinced us that vibecoding scientific software is now viable in record time with limited resources. We proved that even very complex computational physics codebases can be modernized rapidly. 🚀
English
0
1
6
413
Eric Volkmann
Eric Volkmann@e_volkmann·
@ggalletti_ JAX’s native AD enables differentiable physics out of the box. To showcase a use case, we use gradient-based optimizers to solve inverse problems, such as recovering temperature gradients (R/LT​) and performing exact sensitivity analysis on the growth rates.
English
1
1
4
463
Eric Volkmann
Eric Volkmann@e_volkmann·
Introducing gyaradax 🐉: A JAX solver for local flux-tube gyrokinetics with custom CUDA kernels for acceleration. This entire code was vibecoded by @ggalletti_ and me in a month. Validated against GKW (CPU-only Fortran code) with 10x speedups. Details and code in the replies.
English
5
15
69
42.9K
Eric Volkmann
Eric Volkmann@e_volkmann·
Preparing for the poster session at #ICML2025 Drop by at W-504 to discuss GINNs: Geometry-Informed Neural Networks. GINNs lie at the Intersection of Geometry, neural shapes and generative modelling. Looking forward to the discussion! 😃
Eric Volkmann tweet media
English
0
1
6
206
Eric Volkmann
Eric Volkmann@e_volkmann·
Alena and I will be presenting our poster today at #NeurIPS2024 in the afternoon session. East hall, #3710 Excited to discuss and answer questions.
Georgia Koppe@GeorgiaKoppe

Can we learn dynamical systems from short, noisy #fMRI data? Building on advanced frameworks for dynamical systems reconstruction (DSR), we present a novel scalable SSM DSR algorithm for fMRI. With Eric Volkmann, Alena Braendle & #DurstewitzLab Explore: arxiv.org/abs/2411.02949

English
0
0
3
104
Eric Volkmann retweetledi
Arturs Berzins
Arturs Berzins@artuursberzins·
Geometry-Informed Neural Networks are evolving! Beyond faster training and improved shapes, GINNs surprised us with an emergent property – a structured latent space. 🧵
GIF
Johannes Brandstetter@jo_brandstetter

Huge progress by @artuursberzins @AndyRadler @e_volkmann on Geometry-Informed Neural Networks (GINNs)! Faster training, better shapes, and surprising insights from enforcing diversity. 📜: arxiv.org/abs/2402.14009 🖥️: arturs-berzins.github.io/GINN/

English
11
145
985
120.7K

Keşfet

@LL_AIlo @fchollet @ggalletti_ @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates