Computer Vision News - January 2018

Computer Vision News and the form of data storage in the brain, which can be accessed at the appropriate resolution for a given scenario. A great deal of this know-how is stored in the visual and motor cortices, which act for us as an internal representation of our world. Perception, to be efficient, must be able to handle a wide variety of scenarios concurrently. The more specific question -- what type of computer model is sufficient for simulating perception in the human brain? One of the means of approaching the answer to this question is asking: what would a model representing human vision look like? Then try to expand from there to a model for all perception. In this paper, the authors take a step towards answering these questions. They demonstrate how clues and hypotheses from neuroscience about the structure of the visual cortex can be combined to produce a computer vision model. Here is a list of clues and hypotheses the authors used in building the RCN model: Computer Vision News Research 5 Representation in the RCN model Biological observation Surfaces are implemented in the model by Markov Random Field, which enforces continuity, except where interrupted by object contours. Evidence from neuroscience indicates that contour-lines and surfaces are represented in the brain as separate factors. Lateral connections are implemented by pool variables being connected by factors that enforce compatibility between the choices made in different pools. Lateral connections are a predominant feature of the visual cortex. They have been amply observed and documented by research. The model implements top-down object- based attention by a combination of non- negative weights and lateral connections. Research has shown the visual cortex has excellent capacity for distinguishing between object instances, even when they are highly overlapping and/or transparent. This quality is known as top-down object-based attention. Representational choices were beneficial for message-passing-inference: include feature-specific lateral connections and sparsity of weights. Neuroscience evidence shows the visual cortex makes use of message-passing-based approximate inference and learning. Research