Skip to main content

Towards Stronger AI by Out-of-Distribution Testing: CompNets Versus Deep Nets

ECE Colloquia Virtual Seminar

-
Location: Current students will receive Zoom access details via Canvas
Speaker:
Alan Yuille
Johns Hopkins University

AI Vision algorithms appear to be superhuman when evaluated on standard benchmarked datasets, but in reality they are far less robust and general purpose than the human visual system. We argue that evaluating AI using standard benchmarks is misleading because the algorithms take short cuts which  exploit the biases in the datasets and which downplay the corner cases. We argue that in order to overcome the limitations of current AI we need to  evaluate their performance  more toughly by, for example, using adversarial examiners which try to pinpoint the weakness of algorithms and out-of-distribution tests to see if they can generalize outside the training data. We hypothesize that Bayesian generative approaches will perform better than standard SOTA deep networks under these more challenging situations. To illustrate this we describe a research program where we develop simple generative models, which we call CompNets, and show that they can outperform SOTA deep networks on a set of challenging tasks including robustness to heavy occlusion, robustness to patch attacks, performing amodal segmentation, and transferring to different domains.

Yuille

Alan Yuille received his B.A. in mathematics from the University of Cambridge in 1976, and completed his Ph.D. in theoretical physics at Cambridge in 1980, supervised by Stephen Hawking, by postdoctoral work in 1981 at the Physics Department, University of Texas at Austin, and the Institute for Theoretical Physics, Santa Barbara. He became a research scientist at the Artificial Intelligence Laboratory at MIT (1982-1986), followed by postdoctoral research and then a faculty position in the Division of Applied Sciences at Harvard (1986-1995). From 1995-2002 he worked as a senior scientist at the Smith-Kettlewell Eye Research Institute in San Francisco. From 2002-2016 he was a full professor  in the Department of Statistics at UCLA with joint appointments in Psychology, Computer Science, and Psychiatry. In 2016 he became a Bloomberg Distinguished Professor in Cognitive Science and Computer Science at Johns Hopkins University. He has won a Marr prize, a Helmholtz prize, and is a Fellow of IEEE. He has broad research interests in vision, machine learning, cognitive science, and neuroscience. He has over 500 peer reviewed publications, over 75,000 citations, and an h-number of 118.

Seminar Series