Francisco Sahli Costabal

Assistant Professor at @ucatolica | Machine learning & Cardiac Biophysics

[Juan Molina @xJJMM07]

Did you know neural networks have a frequency bias when they learn? Namely, they learn low frequencies before high ones.
We explain this phenomenon and offer tools to eliminate/control it.

➡️ Learn more: https://t.co/Cz9AVM5mf8 🔥

Thread 🧵 (1/7)
@fcosahli @MirceaSci

[Francisco Sahli Costabal @fcosahli]

@ohowell123 @ParisPerdikaris @SimPezzu yes, we only do the eigenfunction computation once and it's pretty fast compared to the training time. We have a Figure in the paper regarding this. The mesh discretization influences the results, but it's not too critical for the low freq. eigenfunctions.

[Francisco Sahli Costabal @fcosahli]

new #preprint:

Δ-PINNs: physics-informed neural networks on complex geometries

We encode the domain using the Laplacian eigenfuctions, then we solve forward and inverse problems on any geometry, like geodesics on a bunny!👇

https://t.co/GQAcX2i7si

@ParisPerdikaris @SimPezzu

[Francisco Sahli Costabal @fcosahli]

New #preprint:

Probabilistic learning of the Purkinje network from the electrocardiogram

We identify potential representations of ❤️the cardiac conduction system❤️that explain patient's electrocardiogram with and without pathologies.

https://t.co/Jk2MqdsCqL

@SimPezzu

[Adrian Buganza Tepole @abuganzat]

🚨New #preprint 🚨
https://t.co/26fzsywAxk

'Generative hyperelasticity with physics-informed diffusion fields'!
with @fcosahli @tajtac @ProfRausch @BilionisIlias

We use diffusion for uncertainty quantification in material models.

A 🧵👇

[Francisco Sahli Costabal @fcosahli]

New #Preprint “Unsupervised reconstruction of cardiac cine MRI using Neural Fields”

We reconstruct dynamic images of the ❤️ with fully connected neural networks using only a fraction of the data, reducing acquisition times ⏳

https://t.co/4J1rR4PRai

#neuralfields #MRI

[Francisco Sahli Costabal @fcosahli]

@mholla_back Cool! What is it for? Beam deflection?

[Adrian Buganza Tepole @abuganzat]

New preprint work from the lab by @tajtac in collaboration with @fcosahli and @ProfRausch: Data-driven anisotropic finite viscoelasticity using neural ODEs https://t.co/PH1e8ap8bO #MachineLearning #DataScience

[Francisco Sahli Costabal @fcosahli]

new #preprint!

WarpPINN: Cine-MR image registration with physics-informed neural networks

We learn the deformation of the♥️using a continuous space-time NN, informed by the mechanics of cardiac tissue.

https://t.co/CrkL144zpq

@ArratiaLo1 @DanielEHurtado1 @hernan_mella

[Francisco Sahli Costabal @fcosahli]

@riceasphait @KoestlerLukas @neural_fields @eccvconf @danielgrittner_ @ZorahLaehner really cool! and is amazing to see very little overlap in the applications of same idea over the 3 papers. Definitively there is huge potential in this area.

[Francisco Sahli Costabal @fcosahli]

@KoestlerLukas @neural_fields @eccvconf @danielgrittner_ @ZorahLaehner Thanks! Will definitely send you a message to discuss more. We also added a reference to your paper in our final submission :)

[Francisco Sahli Costabal @fcosahli]

@keenanisalive @unsorsodicorda @ParisPerdikaris @SimPezzu @NVIDIAAI yes! I have your method implemented and I use it quite a bit.

In this paper we attempt to learn the location of the b.c. from observed distances.

In my experience, the heat method does not satisfy the Eikonal equation as well as the exact geodesics, which may be an issue here.

[Francisco Sahli Costabal @fcosahli]

@unsorsodicorda @ParisPerdikaris @SimPezzu @keenanisalive @NVIDIAAI this is problematic because the operators of the pde are not defined on the surface when you use automatic differentiation in standard PINNs.

5) thanks! I think there are a lot of new things to discover in this area.

[Francisco Sahli Costabal @fcosahli]

@unsorsodicorda @ParisPerdikaris @SimPezzu @keenanisalive @NVIDIAAI 4) thanks for the reference, and the main difference is that we work on surfaces rather than volumes, so you can see it more as a re-parametrization of the domain rather imposing b.c.s. The main issue that we tackle is that the input of the nn is (x,y,z) but the surface is 2D.

[Francisco Sahli Costabal @fcosahli]

Thanks for the invitation!

[Closer Look Journal Club @closerlookjc]

Yesterday, @fcosahli opened the Fall '22 season of the Closer Look 🧐 journal club. Its a beautiful talk on his recent paper (👉https://t.co/NH8cZqet0H), which also includes a neat review on Gaussian Processes! If you missed it, watch it here:
https://t.co/0FO5yI1Jkx

[Closer Look Journal Club @closerlookjc]

Welcome back to school 🎒🏫! We are back with more journal club, join us tomorrow Wednesday 24th at 11:30am ET to learn from @fcosahli about 🫀🩺📈. More info about our Fall lineup: https://t.co/njOWDQnSeM