Join us this Friday, July 10th at 3pm exclusively on zoom for the first meeting of the Stats and ML Journal Club for the quarter. This week, Elise Darragh-Ford will lead a discussion on Scattering transforms for cosmology, a promising alternative to direct machine learning. See you then!
Abstract: Parameter estimation with non-Gaussian stochastic fields is a common challenge in astrophysics and cosmology. In this paper we advocate performing this task using the scattering transform, a statistical tool rooted in the mathematical properties of convolutional neural nets. This estimator can characterize a complex field, without
explicitly computing higher-order statistics, thus avoiding the high variance and dimensionality problems. It generates a compact set of coefficients which can be used as robust summary statistics for non-Gaussian information. It is especially suited for fields presenting localized structures and hierarchical clustering, such as the
cosmological density field. To demonstrate its power, we apply this estimator to the cosmological parameter inference problem in the context of weak lensing. Using simulated convergence maps with realistic noise, the scattering transform outperforms the power spectrum and peak counts, and is on par with the state-of-the-art CNN. It retains the
advantages of traditional statistical descriptors (it does not require any training nor tuning), has provable stability properties, allows to check for systematics, and importantly, the scattering coefficients are interpretable. It is a powerful and attractive estimator for observational cosmology and, in general, the study of physically-motivated fields.