Sequential recognition of superimposed patterns with top-down selective attention

Abstract

Top-down attention is a cognitive mechanism to filter out irrelevant information from sensory input. Unlike bottom-up attention based on the sensory signal itself the top-down attention process is originated from the higher brain, which consists of previous knowledge about the sensory signals. A simple computational model is developed for the top-down attention. In this model an attention gain coefficient is assigned to each input feature, and all the attention gain coefficients are dynamically adjusted based on previous knowledge. A multilayer Perceptron is used to model the knowledge in the higher brain. The developed model demonstrates excellent capability of extracting and recognizing each pattern sequentially from superimposed dual-class patterns studied in visual perception. (C) 2004 Elsevier B.V. All rights reserved.