Date: November 10th, 2021 (Wednesday), 2:00-3:00pm (Central Time)
Talk type: Hybrid – In-person (Peter O’Donell Building, Room # 6.304) and Zoom (link)
Talk Title: Theory and Practice of Generative Autoencoders in Deep Learning
Talk Overview:We will define autoencoders and the various ways they can be used for computer vision tasks. Then we will talk about the class of generative autoencoders, which are autoencoders trained in a way that can be used to generate new data distributed similarly to the data on which it was trained. These are commonly cited as an alternative to the popular GAN framework (Goodfellow et. al., 2014) for generative models in computer vision. This will include a short survey of the most popular architectures for generative autoencoders, including Variational Autoencoders (Kingma-Welling 2014) and Adversarial Autoencoders (Makhzani et. al 2015). Time permitting, we will also demonstrate these things in practice with a short demo.
Speaker Bio: George is a fifth-year PhD candidate in the Mathematics department at UT-Austin. His interests are in algebraic geometry, tropical geometry, machine learning, and optimization. Specifically, he works on computational and enumerative problems associated to moduli spaces of curves. He is also funded by the Institute for Foundations of Machine Learning. Outside of his PhD research, he has spent part of his time over the past few years as a data scientist and machine learning engineer at Striveworks.