A selective attention neural network for object recognition

Abstract

A selective attention neural network for recognizing multiple objects in any position or orientation is proposed. Selective attention controls where the neural network focuses in the scene and at what times. By using selective attention, neurons are traded for speed. Since the network does not process all the inputs in parallel, it is possible to build neural networks that can handle multiple objects and high-resolution scene images using current simulators. An object recognition module has been developed and implemented using a recurrent neural network. The object recognition module takes the output of the selective attention module, builds up an internal representation of the object and makes a decision on the object's identity using the internal representation. Preliminary results have shown that the object recognition network decides on the object's identity before processing the entire object. The network needs to process only part of the object before making its decision. This allows a neural network to recognize partially occluded objects, a difficult problem using classical techniques