Color Encoding in Convolutional Neural Networks Webinar
Convolutional Neural Networks (CNN) have been proposed as suitable engines to solve computer vision problems. Their impressive performance is a bit shadowed by their black-box nature and the consequent lack of understanding of how the visual information is internally represented. This talk shows the results of dissecting one of these networks trained on more than a million of images to perform an object recognition task.
The task focuses on analyzing how color is represented by individual neurons by defining a color selectivity index. We find color opponency clearly comes up in the first layer; in the second layer color selectivity is tuned to a more dense sampling; and in deeper layers the neuros are selective to specific colored patterns, like specific colored objects (e.g., orangish faces), surrounds (e.g., top blue sky) or object-surround configurations (e.g. red blob in a green surround as a ladybug detector). Overall, the work is revealing certain analogies between CNN intermediate representations and evidences reported in studies of color encoding in primate brains.
About the presenter:
Maria Vanrell is associate professor at the Computer Science Department of Universitat Autònoma de Barcelona (UAB). She graduated in computer engineering (1990) and received her PhD (1991) from UAB. She is currently attached to the Computer Vision Centre where she created the Color in Context research group and is head of projects. Her main publications are essentially addressed to color in computer vision with some focus on bio-inspired considerations for computational color models. Main contributions are in the problems of predicting color saliency, naming, texture, induction, constancy, segmentation, recognition, and decomposition in intrinsic components from a single image.