PIRSA:21120020

The scattering transform in cosmology, or, a CNN without training

APA

Cheng, S. (2021). The scattering transform in cosmology, or, a CNN without training. Perimeter Institute. https://pirsa.org/21120020

MLA

Cheng, Sihao. The scattering transform in cosmology, or, a CNN without training. Perimeter Institute, Dec. 13, 2021, https://pirsa.org/21120020

BibTex

          @misc{ pirsa_21120020,
            doi = {10.48660/21120020},
            url = {https://pirsa.org/21120020},
            author = {Cheng, Sihao},
            keywords = {Cosmology},
            language = {en},
            title = {The scattering transform in cosmology, or, a CNN without training},
            publisher = {Perimeter Institute},
            year = {2021},
            month = {dec},
            note = {PIRSA:21120020 see, \url{https://pirsa.org}}
          }
          

Sihao Cheng Johns Hopkins University

Abstract

Patterns and complex textures are ubiquitous in astronomical data but challenging to quantify. I will present a new powerful statistic called the “scattering transform”. It borrows ideas from convolutional neural nets (CNNs) while retaining the advantages of traditional statistics. As an example, I will show its application to weak lensing cosmology, where it outperforms classic statistics and is on a par with CNNs. I will also show interesting visual interpretations of the scattering statistics and possible extensions of this “mathematical neural network” idea. I argue that the scattering transform provides a powerful new approach in cosmology and beyond.

Related papers:

https://arxiv.org/abs/2112.01288

https://arxiv.org/abs/2103.09247

https://arxiv.org/abs/2006.08561



Zoom Link: https://pitp.zoom.us/j/91612161747?pwd=bnQrVmo4ZjBjaUdQMDBNZGhFS2NPQT09