SDSC-AI4Science seminar: Advancing cosmology with deep learning
Event details
| Date | 22.03.2023 |
| Hour | 16:00 › 17:00 |
| Speaker | Dr. Tomasz Kacprzak and Dr. Nathanaël Perraudin, Sr. Data Scientists at SDSC |
| Location | Online |
| Category | Conferences - Seminars |
| Event Language | English |
Place: GA3 21 (Bernoulli Center), followed by an apéro in GA3 31. (Zoom: https://epfl.zoom.us/j/67864899458)
Speaker: Dr. Tomasz Kacprzak and Dr. Nathanaël Perraudin, Sr. Data Scientists at SDSC
Title: Advancing cosmology with deep learning
Abstract: Cosmology, despite its seemingly disparate nature from machine learning, is actually a data-driven field that can greatly benefit from recent advancements in machine learning techniques. In this presentation, we will showcase a collaborative SDSC project that exemplifies how machine learning can contribute to significant breakthroughs in cosmology.
In the first part of the talk, we will give an overview the types of innovation that we applied to cosmology: cosmological parameter measurements from dark matter maps using deep learning, accelerating cosmological simulations using Generative Adversarial Networks, and analysing data on the sphere using graph convolutional neural networks.
In the second part of the talk, we will do a deep-dive into the problem of machine learning on the sphere using graphs and present our approach called DeepSphere. Graphs are versatile data structures that can represent pairwise relationships between objects or act as a discrete representation of a continuous manifold. Using the graph-based representation, we define many of the standard CNN operations, such as convolution and pooling. With filters restricted to having radial symmetry, our convolutions are equivariant to rotation on the sphere, and DeepSphere can be made invariant or equivariant to rotation. This way, DeepSphere is a special case of a graph CNN, tailored to the HEALPix sampling of the sphere. This approach is computationally more efficient than using spherical harmonics to perform convolutions. We demonstrate the method on a classification problem of weak lensing mass maps from two cosmological models and compare its performance with that of three baseline classifiers, two based on the power spectrum and pixel density histogram, and a classical 2D CNN. Our experimental results show that the performance of DeepSphere is always superior or equal to the baselines.
This research project was performed in close collaboration with the Cosmology Research Group of ETH Zürich.
Speaker: Dr. Tomasz Kacprzak and Dr. Nathanaël Perraudin, Sr. Data Scientists at SDSC
Title: Advancing cosmology with deep learning
Abstract: Cosmology, despite its seemingly disparate nature from machine learning, is actually a data-driven field that can greatly benefit from recent advancements in machine learning techniques. In this presentation, we will showcase a collaborative SDSC project that exemplifies how machine learning can contribute to significant breakthroughs in cosmology.
In the first part of the talk, we will give an overview the types of innovation that we applied to cosmology: cosmological parameter measurements from dark matter maps using deep learning, accelerating cosmological simulations using Generative Adversarial Networks, and analysing data on the sphere using graph convolutional neural networks.
In the second part of the talk, we will do a deep-dive into the problem of machine learning on the sphere using graphs and present our approach called DeepSphere. Graphs are versatile data structures that can represent pairwise relationships between objects or act as a discrete representation of a continuous manifold. Using the graph-based representation, we define many of the standard CNN operations, such as convolution and pooling. With filters restricted to having radial symmetry, our convolutions are equivariant to rotation on the sphere, and DeepSphere can be made invariant or equivariant to rotation. This way, DeepSphere is a special case of a graph CNN, tailored to the HEALPix sampling of the sphere. This approach is computationally more efficient than using spherical harmonics to perform convolutions. We demonstrate the method on a classification problem of weak lensing mass maps from two cosmological models and compare its performance with that of three baseline classifiers, two based on the power spectrum and pixel density histogram, and a classical 2D CNN. Our experimental results show that the performance of DeepSphere is always superior or equal to the baselines.
This research project was performed in close collaboration with the Cosmology Research Group of ETH Zürich.
Practical information
- Informed public
- Free
- This event is internal