Neural network modelling of the primate ventral visual pathway

The aim of this doctoral research is to advance understanding of how the primate brain learns to process the detailed spatial form of natural visual scenes. Neurons in successive stages of the primate ventral visual pathway encode the spatial structure of visual objects and faces. However, it remains a difficult challenge to understand exactly how these neurons develop their response properties through visually guided learning. This thesis approaches this problem through the use of computational modelling. In particular, I first show how the brain may learn to represent the spatial structure of objects and faces through a series of processing stages along the ventral visual pathway. Then I propose how understanding the two complementary unsupervised learning mechanisms of translation invariance may have useful applications in clinical psychology. Next, the potential functional role of top-down (feedback) propagation of visual information in the brain in driving the development of border ownership cells, which are thought to play a role in binding visual features such as boundary edges to their respective objects, is investigated. In particular, the limitations of traditional rate-coded neural networks in modelling these cells are identified. Finally, a general solution to such binding problems with the use of a more biologically realistic spiking neural network is presented. This work is set to make an important contribution towards understanding how the visual system learns to encode the detailed spatial structure of objects and faces within scenes, including representing the binding relations between the visual features that comprise those objects and faces.