Self-supervised Learning with Physics-aware Neural Networks I: Galaxy Model Fitting.

This paper presents a very simple but powerful idea. Instead of supervising a neural network during training we give it a model for reconstructing the input from a set of predicted parameters. This architecture, based on the autoencoder, has many applications where we do not have access todo the initial conditions of the problem or for when we don't even know how to generate correct input images from a model.

Read the paper

Classifying the Large Scale Structure of the Universe with Deep Neural Networks.

What do cosmological filaments and walls have in common with soft tissue and bone? it is more than 10 years ago that I developed the Multiscale Morphology Filter (MMF) based on medical imaging techniques for vessel and bone segmentation using Hessian filters. This papers presents the first application of Deep Convolutional Neural Networks to the segmentation of cosmological filaments and walls. The main idea (U-Net) was developed for medical imaging problems.

The website with extra figures will be updated on 4th April.

Predicting cosmic environment with machine learning.

We are trying to predict the cosmic environment of a galaxy based on the properties of the galaxy. This is essencially the inverse of what is traditionally done in galaxy-environment studies where galaxy properties are predicted from environment. What we learned is that the assembly history of a galaxy is sensitive to its environment so simple criteria to select galaxies like the Milky Way by mass and assembly history make sense after all.

Read the paper