An analog VLSI implementation of an attention-based saccade generator /

A huge amount of data is present in the visual field of the primate visual system. Serial processing of the entire visual field requires long periods of time, while parallel processing involves very complex architectures. Biological systems are able to handle this problem using attention. Attention is a mechanism where interesting locations in a given scene, that are worthy of more processing, are identified. These locations are then serially attended to where the precious amount of resources are allocated for further parallel processing. Biological systems have been able to find a good compromise between the slow serial architecture and the complex parallel one. In this thesis, we present an analog VLSI implementation of a system that mimics the attentional mechanism of the primate visual system. We follow the premotor theory of attention in that attention and motor action (specifically saccadic eye movements) are tightly linked. These two processes have been considered separately in most previous computational and VLSI models of vision. Our model facilitates the selection of regions of interest by using a winner-take-all mechanism. This defines a set of inhibitory signals, one at each node, that determine the behaviour of the system. It is the pattern of these inhibitory signals that gives rise to saccadic eye movements. The model also contains an inhibition-of-return mechanism, a proximity preference behaviour, and a centroid detection circuit. One and two-dimensional versions of the model were fabricated and tested. The use of the analog VLSI paradigm to implement our model has provided a powerful mechanism to model the massively parallel architecture of biological systems