Extracting information from stochastic fields is a ubiquitous task in science. However, from cosmology to biology, it tends to be done either through a power spectrum analysis, which is often too limited, or through the use of convolutional neural networks, which require large training sets and lack interpretability.
I will present a powerful statistical tool called the “scattering transform”, which stands nicely between the two extremes, and recent updates to extend this idea. I will use various examples in cosmology and beyond, including its recent application to HSC weak lensing data, to demonstrate its efficiency, interpretability, and advantage over traditional statistics.
Bio:
Sihao Cheng is a postdoc member of the Institute for Advanced Study and a visiting fellow of the Perimeter Institute. He obtained his Ph.D. from Johns Hopkins University and Bachelor degree from Peking University. He uses innovative and interdisciplinary ideas to analyze survey data and acquire new physical understandings. His work led to the discovery of special stars powered by gravitational energy while they are crystallizing and the cosmological applications of a new statistic that borrows ideas from deep learning. His interest spans from cosmology to (exo)planets.