Stephan Mandt, an assistant professor of computer science and statistics in UCI’s Donald Bren School of Information and Computer Sciences (ICS), received an unrestricted gift of $50,000 from Disney Research Los Angeles to support his work on generative modeling.

Mandt is exploring how to use machine learning under resource constraints — in particular, given a lack of relevant training data. “In other words, how might you train a deep learning algorithm to mimic speech when you don’t have massive amounts of available speech recordings?”

Mandt is currently teaching a related course (CS 295: Deep Generative Models), examining how deep learning models can generate new data such as video, text, speech or images. These models generate artificial data that humans cannot easily distinguish from real-world data. At the same time, deep generative models can also be used to compress data such as images with unprecedented performance.

“This course is essentially along the lines of some of my work on generating and compressing images and video,” he says. “We can apply similar tools that we use to generate artificial video and other sequential data to generate acoustic data and speech.”

Mandt’s approach toward resource-efficient deep learning uses a unified set of mathematical and statistical methods, where he combines ideas from deep learning, Bayesian statistics, and information theory. “Making deep learning more resource-efficient is of enormous societal relevance,” he explained after receiving an NSF CAREER Award earlier this year. “Algorithms will make better decisions with less data, making them more reliable in safety-critical areas such as autonomous driving. Also, more data can be processed with less energy and storage.” This unrestricted gift will help advance Mandt’s research and teaching in the area of deep generative modeling.

— Shani Murray