Machine Learning Seminar
Speaker: Chris Adami, MSU
Title: Is It Time to Abandon Deep Convolutional Networks?
Refreshments available at 2:00 pm.
Date: Mon, 05 Feb 2018, 2:15 pm – 3:15 pm
Location: 1400 BPS Bldg.
Deep convolutional neural networks (CNNs) have enjoyed spectacular success in classifying images with complex scenes, but at the same time have been plagued with an Achilles heel that makes them vulnerable to adversarial perturbations. Some theoretical work tracks this vulnerability to the continuous variable nature of the ANN substrate, arguing that digital “firing neuron” implementations could be immune to the dependence on surface statistical regularities. Here I discuss a “digital brain” framework in which digital neurons either fire or not (Markov brains), and show how an evolutionary approach can generate digital brains that robustly classify hand-written digits using just a few “spot-checks” of the image, much like the vertebrate eye saccades over an image to classify it using internal stored representations to fill in the details. Using this approach, it should be possible to not only classify numerals, but also to generate a numeral that is consistent with the acquired measurements. While the evolved digital brains have not been tested for their response to adversarial perturbations, the evolutionary approach strongly suggests that adversarially perturbed images will not be mis-classified.
Dr. Adami obtained his Ph.D. and M.A. in theoretical physics from the State University of New York at Stony Brook, as well as a Diplom in Physics from Bonn University (Germany). His main research focus is Darwinian evolution, which he studies at different levels of organization (from simple molecules to brains). He has pioneered the application of methods from information theory to the study of evolution, and designed the “Avida” system that launched the use of digital life as a tool for investigating basic questions in evolutionary biology. When not overly distracted by living things, he studies the fate of classical and quantum information in black holes. He wrote the textbook “Introduction to Artificial Life” (Springer, 1998), and is the recipient of NASA's Exceptional Achievement Medal. He was elected a fellow of the American Association for the Advancement of Science (AAAS) and a fellow of the American Physical Society (APS).