aboucaud / euclid-school-2023

Content for ML lecture at Euclid-Rodolphe Cledassou summer school 2023

ML lectures - Rodolphe Clédassou summer school 2023

Marc Huertas-Company (IAC) and Alexandre Boucaud (APC)
August 2023

Lectures

Cycle 1

• Introduction to machine learning with (probabilistic) neural networks

Cycle 2

• Introduction and neural networks recap from cycle 1 - HTML slides
• Probabilistic neural networks - PDF
• Convolutional networks - PDF
• Image2image networks and Transformers - PDF
• Attention mechanism and Graph networks - PDF
• Introduction to MLOps - HTML slides

Cycle 3

• Slides to come soon
• Friday Zoom recording | Recording code: CqBP?5b!Y^

Notebooks

Cycle 1

Setup

To run the notebooks locally, install the dependencies from the requirements.txt

python -m pip install -r requirements.txt

Warning

macOS users with M1/M2 processors please follow the instructions below to install TensorFlow (otherwise the notebook kernel will die at the beginning) Apple M1/M2 specific TensorFlow installation

• Neural regression with classic and probabilistic neural networks
• Same notebook with MLFlow examples

Cycle 2

Warning

For cycle 2 and 3, those not using Google Colab links must first run the dataset creation steps below before starting with the notebooks.

Instructions for the notebook:

1. choose one simulation between IllustrisTNG (dataset version 1.0.0) and SIMBA (dataset version 1.0.1)
2. execute Part 1 and 2 whose goal is to predict $\Omega_M$ and try to improve the results of the MLP
3. try to apply the networks trained with one simulation to the other one
4. move on to Part 3 where we try to predict $\sigma_8$ and $\Omega_M$

An alternative is to try the notebook on Google Colab (need a Google account).

Cycle 3

Warning

For cycle 2 and 3, those not using Google Colab links must first run the dataset creation steps below before starting with the notebooks.

References

• https://arxiv.org/abs/2201.02202
• https://camels.readthedocs.io/en/latest/index.html