Our project aims to develop a general architecture for perceptual attention and learning and to implement this architecture in a hardware system based on analogue VLSI.

The perceptual problem to be solved with the help of attention and learning is the classification of dynamic perceptual objects characterized by time evolution (e.g., optical flow, human speech). We will superpose multiple objects in space (vision) or in spectrum (audition). The task of the system will be to attentively track one particular object and to learn to classify it.

The proposed system comprises three parts: feature space, saliency network and associative network. The feature spaces is specific to the stimulus ensemble and the task, and will be designed to represent task-relevant information efficiently and sparsely. It will be implemented in software. In contrast, the saliency and associative networks will be implemented both in software and hardware. The saliency network provides for competitive interactions between different features, allowing one set of features to emerge as a ‘winner’ and other features to be ignored. In others words, the saliency network will provide the selective functionality of attention. The associative network will memorize the distinguishing features of task-relevant objects and thus will learn to classify perceptual objects. Together, the two networks create an opportunity for the virtuous cycle familiar from human perception: attention accelerates learning and learning, once having taken place, can guide attention.