- Location
- Rooms 9014 & 9016, U of T 9th Floor, 700 University Avenue, Toronto, ON M5G 1X6
- Series/Type
- Seminar, U of T Community Event
- Format
- Hybrid
- Dates
- September 25, 2023 from 3:30pm to 4:30pm
Links
Join us at the Statistical Sciences Applied Research and Education Seminar (ARES) with:
Stephen Portillo
Assistant Professor of Physics
Concordia University of Edmonton
Free Hybrid (In-person/Online) Event | Registration Required
Talk Title
From Pixels to Parameters
Abstract
Much of astronomy uses pixelized data, but the size and complexity of these data often strain the capability of existing data analysis techniques. I will present algorithms built on advances in statistics and machine learning that allow more science to be done with the same pixels. Digital tracking searches for Kuiper belt objects (KBOs) involve series of images where the KBO is undetectable in each image, but detectable in the series. By forward modelling the position of the KBO in each image in a “joint-fit”, the KBOs’ trajectories can be measured precisely enough to constrain their dynamics. Probabilistic cataloguing (PCat) is a reversible jump Markov chain Monte Carlo method that creates model images for an unknown number of sources. In extremely crowded fields with stars every 10 pixels, PCat finds stars four times fainter than DAOPHOT, a commonly used pipeline. Finally, I will discuss a variational autoencoder, a type of deep generative model, that I have trained on galaxy spectra from the Sloan Digital Sky Survey. This autoencoder learns a six dimensional latent space that naturally separates different classes of galaxy and captures variation in spectral line widths and ratios.
Speaker Profile
Dr. Portillo is an Assistant Professor of Physics at Concordia University of Edmonton. He obtained his PhD in Astronomy and Astrophysics from Harvard University and was a DIRAC Postdoctoral Fellow at the University of Washington. His research applies advances in statistics and machine learning to enable more science to be done with astronomical data sets. The techniques he uses in his research include Bayesian inference, reversible jump Markov chain Monte Carlo, and variational autoencoders.
