facusapienza21 / dynamical-lasso

Hello, I'm Facundo (Facu) Sapienza 👋!

I am a PhD candidate in the Department of Statistics at the University of California Berkeley working under the supervision of Fernando Pérez and Jonathan Taylor. Previourly finished my Bsc/Msc (Licenciatura) in Physics and Mathematics in the University of Buenos Aires🇦🇷.

• 🔭 I’m currently working on:
  • 🧊 ODINN.jl: Global glacier model using Universal Differential Equations for climate-glacier interactions. Integrating the Julia differential programming tools with the Open Global Glacier Model (OGGM) in Python.
  • 🌎 PolarWandering: Statistical modelling of Paleomagnetic data.
• 🌱 I’m currently learning about sensitivity methods for differential equations. Interested in participating? Check out our review paper and contribute to it!
• 👨‍🏫 I am interested in teaching machine learning for geophysisc and good practices fordeveloping open, reproducible and collaborative scientific software.
  • Collaborative and Reproducible Data Science at UC Berkeley: Course given to undergraduate and graduate students.
  • Machine Learning in Glaciology Workshop: Given in Finse, Norway, to students and researchers working in the intersection of cryospheric sciences 🧊 and machine learning 🤖.
• 👯 I’m looking to collaborate with people interested in the intersection between machine learning, geophysics and scientific computing.
• 📫 How to reach me: feel free to send me an email to fsapienza@berkeley@edu.
• 😄 Pronouns: He/Him

dynamical-lasso's Issues

Double-step optimization to find age uncertainties

How to deal with non-parametric problems where we are uncertain about the age? We can replace the $t_i$ with $t_i + \eta_i$ where the $\eta_i$ are enforced to be small (and probably sparse) with a Huber/ElasticNet penalization term, but where we can optimize on them as the same time we find the UDE.

Make AD work with `Lux.jl`

The default now for SciML is moving from Flux.jl to Lux.jl. This means that the example of the UDE needs to be adapted to work with Lux.jl, which requieres to convert the parameter object to a parameter type for lux so that the AD machinery works. This is explained in this Julia Discurse post and addresed in this example.

Explore non-parametric regression using UDE in a generic case

In principle, this method serves as a way of doing non-parametric regression in the sphere, since any path can be approximated as the solution of the UDE. It is up to us which constraints we want to add, for example, do we want smooth rotations? Sparsity? Etc. This will require to find ways to efficiently evaluate derivatives of NN with respect to input parameters, which for some reason is not by default fast in with Lux.