Attention Please!
Ernst Niebur, Ph.D.,
Johns Hopkins University
Abstract: To cope with the high complexity of cluttered visual scenes in the presence of limited processing capacity, biological as well as engineered visual systems need to deploy sophisticated strategies. We will mainly discuss data-driven (bottom-up) strategies, based on saliency map algorithms, and also touch on some goal-driven (top-down) strategies. To further increase computational efficiency, primates use attention to objects rather than to elementary visual features. This requires the organization of the visual scene into proto-objects, the pre-cursors of perceptual objects with very simple properties. Organization of the scene into proto-objects transforms the seemingly impossible task of scene understanding into manageable sub-tasks. Attentional selection can work on proto-objects, and tasks like object recognition can then proceed in a sequential fashion, by operating on one proto-object at a time. We propose that primates solve this perceptual organization task using small populations of dedicated neurons that represent different (proto-)objects, and we propose that this is also an efficient structure in computational vision.
ORGANIZERS
David C. Johnson, The Graduate Center and York College, CUNY
Dina Lipkind, The Graduate Center and York College, CUNY