Part 1:
Title: Encoding of Spectra and Time Series
Speaker: Peter Melchior (Assistant Professor of Statistical Astronomy, Princeton University)

Part 2:
Title: Strategies for variance reduction in spectral unmixing
Speaker: Jordan Bryan (Assistant Professor of Data Science, UVA)

Venue: NRAO (ER-230) Pizza will be served!

Zoom https://utexas.zoom.us/j/87159746528?pwd=ouQu8lN9ARbb6aRFpvdf6Ddb1Oqa8B.1
Meeting ID: 871 5974 6528 | Passcode: 996082
Join by SIP: 87159746528@zoomcrc.com 

Part 1 Abstract: Ongoing and future surveys will surprise us, but only if our methods have the speed and flexibility to take advantage of vast quantities of data. I will present a neural network architecture that is specifically designed for galaxy spectra at variable redshifts. It is trained with an invariance-promoting loss function and can create superresolution reconstructions and infer the physical properties of galaxies with few-shot learning. The representations of optical spectra provide accurate prediction of IR photometry, a connection that is not captured by current physical spectrum modeling methods. I will discuss extensions of this work to exoplanet and quasar spectra to demonstrate the strengths and versatility of this approach.

A very similar encoder-decoder architecture can be taken for the modeling of time series, but I will add another component: a Latent ODE model describes temporal evolution as the solution of a differential equation, which is discovered from the data. I will demonstrate the use of this method for exoplanet detection and what it reveals about the training dynamics of deep neural networks.

Part 2 Abstract: Spectral data arise in many scientific fields including biology, environmental monitoring, and astronomy. Physical laws and simplifying statistical assumptions motivate the idea that a mixed spectrum is approximately a linear combination of several constituent spectra. In the problems we consider, the constituent spectra are derived from chemical components with known labels, while the non-negative weights of the linear combination represent the abundance of these chemicals in the mixture. In this context, we propose strategies for using a dictionary of labeled spectra to estimate the abundance weights. The estimates we propose leverage statistical characteristics of spectral data in order to achieve lower variance than standard alternatives. We illustrate the use of these estimates on fluorescence spectroscopy measurements for water quality monitoring. Finally, we discuss the limitations of our approach as well as outstanding statistical questions, particularly as they relate to applications in astronomy such as quasar absorption line spectroscopy.

Bios:

Prof. Peter Melchior leads the Princeton Astro Data Lab, which develops new algorithms to solve problems that hold back astronomy. His group seeks to optimally extract information about astrophysical processes with novel techniques that combine physics principles with deep learning. He creates methods for signal separation, data fusion, fast inference, and outlier detection for large astronomical surveys, and is funded by NSF, NASA, the Keck Foundation, and the Schmidt Sciences Foundation.

Jordan Bryan is a statistician with broad expertise in multivariate data analysis. He has studied and developed statistical methods in the fields of environmental monitoring, high-energy physics, and cancer genomics. His research interests include Bayesian statistics, robust estimation, and information-assisted hypothesis testing.

Prior to joining the faculty at UVA, he was a postdoctoral researcher at the University of North Carolina at Chapel Hill, where he was supported by training grants from the National Institute of Environmental Health Sciences (NIEHS) and the National Heart Lung and Blood Institute (NHLBI). He also worked as an associate computational biologist at the Broad Institute of MIT and Harvard. As of January 2024, he is the Secretary of the junior section of the International Society for Bayesian Analysis. He received his Ph.D. in Statistics from Duke University in 2023.