Explore deep learning insights into cosmological structure formation in this 50-minute talk by Luisa Lucie-Smith from the Institute for Advanced Study. Delve into the fundamental role of dark matter halos in cosmic large-scale structure and their impact on galaxy formation. Learn about a 3D convolutional neural network (CNN) model trained to predict dark matter halo mass from initial conditions density fields in N-body simulations. Discover a technique for interpreting the CNN's learned features in relation to physical properties of the Universe's initial conditions. Gain understanding of machine learning applications in astronomy, cosmological structure formation, and dark matter halo collapse. Examine the construction of initial conditions-to-halo mass mapping, the rationale behind using CNNs, and the training process involving loss function minimization. Explore ongoing work in knowledge extraction from neural networks and assess the CNN's ability to extract relevant features for halo mass predictions.
Deep Learning Insights into Cosmological Structure Formation - Luisa Lucie-Smith
Institute for Advanced Study via YouTube
Overview
Syllabus
Intro
Machine learning in Astronomy
Cosmological structure formation
Dark matter halo formation
Insights into dark matter halo collapse from ML?
A deep learning (DL) approach to halo formation
Construct ICs-to-halo mass mapping for every particle
Why convolutional neural networks?
CNN model
Convolutional layer
Training the CNN: minimising a loss function
Interpreting the learnt mapping between ICs and halos
Extracting knowledge from a neural network
Work in progress: knowledge extraction
Conclusions
Testing the ability of the CNN to extract features
Halo mass predictions from the initial conditions
Taught by
Institute for Advanced Study
