The School of Informatics, Computing, and Engineering (SICE) CS Colloquium Series
Speaker: Etienne Barnard, Research Professor, North-West University, Hermanus, South Africa
When: 1:15 pm, Friday, August 31, 2018
Where: Luddy Hall, Rm. 1106 (Dorsey Learning Hall)
Topic: Generalization in Deep Learning: who buys lunch?
Abstract: The successes that have been achieved with Deep Neural Nets in recent years rely on the unprecedented ability of these networks to generalize on unseen data. However, this property of DNNs is somewhat mysterious, since it occurs in models with many more parameters than conventional machine-learning theories would recommend. I review the empirical evidence for such paradoxical generalization in DNNs, and then present a few perspectives on how generalization may arise under these circumstances. These approaches (which include the controversial "no free lunch theorem") do not explain the performance of DNNs, but provide some potentially useful insights. I will end with speculations on a more satisfactory basis for understanding generalization.
Bio Summary: Etienne Barnard obtained his Ph. D. from Carnegie Mellon University in 1989, and has since then been active in research and development in pattern recognition and speech processing. Barnard has co-authored more than 250 refereed scientific publications on neural networks and speech recognition. He holds a number of international patents and is a past Associate Editor of the IEEE Transactions on Neural Networks. Barnard’s research contributions have been recognized in various ways, including a number of Google Research Awards; he is currently Research Professor at the North-West University based in Hermanus, South Africa.