How do humans reason about and find meaning in images? Despite advances in artificial neural networks such as deep neural networks for computer vision applications, extracting salient information from images and then reasoning about it in the myriad ways that people can remains a challenge. Furthermore, most neural networks tend to be poor scientific models of the human visual system in that they are inconsistent with almost all empirical data in the domain of human vision.
In this talk I will describe a new computational model of human primary visual cortex (visual area V1) that represents the basic elements of shape in the same way that the human visual architecture does. That is, rather than relying on a general purpose architecture that learns statistics derived from extensive training sets, the model uses known properties of the primate cognitive architecture as the basis of shape detection. The model requires no training, has just one free parameter, and does a good job predicting human eye movements, performing well on the Tbingen/MIT salience benchmarks. The model was developed to serve as the front end to a full computational model of human visual reasoning, which is the subject of ongoing work.
I will discuss predictions the model makes about human behavior, practical applications for the geospatial intelligence community, and the benefits that come with emphasis on fidelity to the human cognitive architecture when developing neural networks.