Autonomous attentional selection in SCAN

SCAN (signal channeling attentional network) is a neural network model which realizes a spatial attentional mechanism. The model scans a large input array in order to find a known (sub)pattern. It was previously shown how SCAN is capable of actively shifting attention over the input plane. In the present work, the authors elaborate the original architecture by using Ising-type interactions with a local matching function. The modification provides a local basis for autonomous global attentional shifts toward patterns either expected or salient. It is concluded that the synaptic lattice incorporated in a shifter circuit is capable of rapidly selecting an appropriate input. As part of the SCAN model, its beam can be controlled externally by nonspecific clamping or internally by generating an appropriate (learned) expectation