Neuromorphic implementation of retinotopic arrays of orientation selective hypercolumns

Neuromorphic vision systems inspired by biological systems have advantages of good power efficiency, low hardware cost and small size when compared to digital computer processing, especially since visual processing benefits from parallel computation. The goal of this Ph.D research is to implement a neuromorphic model of the orientation selective hypercolumns found in the visual cortex. We focus upon orientation because it is a basic image feature that is required by higher-level visual processing, such as binocular disparity, motion direction detection and face and pattern recognition. The first part of this thesis describes a new chip architecture for orientation selective image filtering that 1) uses an ON-OFF signal representation inspired by biology to reduce quiescent power consumption by five times and fixed pattern noise by around 22 to 32 times; 2) includes address-event representation (AER) circuits for input and output, which enable the construction of large multichip systems. Chips containing both one-dimensional (1-D) and two-dimensional (2-D) arrays of processing elements have been fabricated and tested. The second part of the thesis describes the use of this chip in constructing the first neuromorphic model of retinotopic arrays of orientation selective hypercolumns. The system contains multiple orientation selective filter chips as well as a silicon retina, which provides input to the filters. Because all of the chips operate in continuous time, this architecture enables us to examine both feed-forward and feedback models of orientation selectivity. This system also opens the door for the development of more complex neuromorphic systems based on the structure and physiology of the visual cortex. For example, it has already been used to construct electronic neurons tuned to binocular disparity, and in a binocular vergence control system