Greg Ver Steeg @ eHarmony
Details
Deep Learning for Insights
12:00 arrival and lunch served
12:30 talk starts
13:30 end of Q&A
Title: Deep Learning for Insights
Please park below 2401 Colorado Ave. and bring your tickets for validation.
Abstract
Existing methods for deep learning work well for structured problems with lots of data like finding cats in images. Most real world data is messy and has unknown value. We introduce a principled and practical new approach to unsupervised learning of deep representations. The approach finds representations that are as informative about the data as possible. The result provides a combination of hierarchical clustering and dimensionality reduction for which the information value of each layer and node in the representation can be quantified. We demonstrate the approach on diverse data from human behavior, language, gene expression, and finance. A preliminary implementation is available at http://github.com/gregversteeg/CorEx .
Bio
Greg Ver Steeg is a research professor in computer science at USC’s Information Sciences Institute. He received his PhD in physics from Caltech in 2009 and since then has focused on using ideas from information theory to understand complex systems like human behavior, biology, and language. His work has been recognized with an AFOSR Young Investigator Award.