A Pragmatic Model of Attention and Anticipation for Active Sensor Systems

Mammals have acquired highly efficient mechanisms for information processing that facilitate quick and adequate responses to environmental changes, and high robustness against distracting signals. Selection of attention is such a fundamental mechanism that leads to a prioritization of sensor information analysis and actuator information synthesis tasks. Different models for selection of attention try to explain this mechanism through the integration of, or interaction between top-down and bottom-up signals and processes. How exactly this integration or interaction occurs in mammals’ brain is still an open problem. We propose in this work a pragmatic model of attention that combines bottom-up (data-driven) analysis with top-down controls, at different levels of sensor and actuator signal processing. The aim of this model is not to explain how attention works in biological systems, but rather to provide a practical design for artificial systems that mimic the behavior of their biological archetype. The main functional elements of this pragmatic model are the selection and fixation of attention, which result from the optimization of an objective function that includes both data-related and knowledge or task-related quantities. The working of the model is illustrated through a generic architecture for complex computer vision tasks, and demonstrated on two applications for detecting and recognizing multiple patterns in complex visual scenes.