New Methods for Learning Deep Visual Hierarchies

Abstract

While humans have an impressive ability in recognizing objects even under adverse observation conditions, computer vision is still constrained to the recognition of specific objects under controlled conditions. In order to reproduce human abilities, visual neuroscientists have long been uncovering important information about the underlying mechanisms that brain may use to perform visual tasks. Nowadays, the use of feedforward visual hierarchies consisting of many layers of linear and nonlinear biologically-inspired operations has shown to give computers an outstanding ability in recognizing faces, animals, vehicles, and a number of other objects. An interesting property of these hierarchies is that they operate directly in the image domain, and that, with simple operations, they are able to learn image processing steps that would instead have to be manually engineered in traditional computer vision approaches. In spite of the recent results, many questions remain about how to learn deep visual hierarchies. Its architecture, for example, was found to be especially relevant. Another relevant aspect is how to train the hierarchy so that it can effectively learn to extract meaningful information from images. In this project, we propose reasonable directions to advance on these aspects. We believe that we have means to address these directions -- presenting preliminary results for one of them -- and that their accomplishment will enable us to achieve new levels of performance in difficult computer vision problems such as unconstrained face recognition, object recognition in general, and content-based image retrieval, among others.