Unsupervised Multimodal Signal Processing with Identity Vectors: A Case Study with Electroencephalograms
Dr. Christian Ward
Wednesday, June 26
12:00 – 1:00pm
1275 Carlson Building
The push for Big Data reaches every field of science. Deriving meaning from this data typically requires labeled data, via experts' domain knowledge, to establish sufficient models for advanced data analysis. However, this is not feasible within the realm of Electroencephalograms (EEGs) given the constraints placed upon clinicians and the amount of data produced during EEG recording sessions. Therefore Identity Vectors, developed with the speech community, have been adapted for use with EEG data as a multimodal signal processing technique. This technique imposes dimensional constraints on the data via a Universal Background Model comprise of Gaussian Mixture Models derived from unlabeled data. Using this as a foundation, the Universal Background Model and subsequently built Total Variability Matrix enable the Identity Vectors to perform competitive classification to other techniques. In addition, this modeling process has been made transparent to the point of highlighting natural modes and relationships in the data.
Dr. Ward's primary research interest is in understanding the statistical connections that drive data classification. He received his PhD in Electrical Engineering from Temple University in May 2019, building on an M.Eng. in Electrical Engineering and a B.S. in Bioengineering from Lehigh University. In addition to data analytics, Dr. Ward has extensive experience implementing and deploying sub-orbital instrumentation based on his work with Orbital Sciences Corporation as an Electrical Engineer supporting NASA’s Sounding Rocket Contract.
When and Where
12:00 PM-1:00 PM
Chester F. Carlson Center for Imaging Science
Open to the Public