Machine learning and statistics: bringing them together for astronomy and cosmology (via Zoom: https://mppmu.zoom.us/j/99300254726)

by Dr Tom Charnock (Institut d'Astrophysique de Paris)

Thursday 4 Feb 2021, 11:00 → 12:00 UTC

Zoom (Main)

Whilst modern machine learning methods, such as deep learning and variational inference, are becoming used wide-spread throughout cosmology and astronomy, there is a fundamental dichotomy between acceptable models describing science and the achievable goals of machine learning. To attempt to address this problem and enable the computational advancements that we see are possible from machine learning, I will describe some of my current research into the statistical interpretation of such methods. By understanding neural networks as a building block of statistical models I will show why they are so adept for tackling current physical problems, but I will also highlight their limitations in terms of comprehensibility and scientific understanding and safety (work based on my Artificial Intelligence for Particle Physics book chapter - Bayesian neural networks (https://arxiv.org/abs/2006.01490 and https://medium.com/@tom_14692/all-deep-learning-is-statistical-model-building-fc310328f07). From this perspective I will highlight a few methods I have been developing to sidestep problems coming from machine learning, whilst still using the machinery. In particular I will show how we can build physically motivated hierarchical models for the inference of cosmology from large scale galaxy surveys (https://arxiv.org/abs/1909.06379 and https://www.aquila-consortium.org/method/machine%20learning/npe.html), how to extract information about physical model parameters (https://arxiv.org/abs/1802.03537) of galaxy morphology from deep survey images and how we can increase the rate of observed rare objects, such as supermassive black holes or high redshift galaxies, by targeting their most likely environments in photometric surveys.