Computational Perceptual Attention (2001)
Micheal Scott Hewett
This dissertation describes CPA, a general-purpose mechanism for expressing and implementing attention policies that control the allocation of resources among sensory processing tasks in a robot or other advanced intelligent system. A wide variety of attention policies can be expressed in this mechanism, which also supports soft real-time constraints on perceptual processing. Intelligent systems can become inundated with data, resulting in perceptual overload and a consequent inability to formulate a timely or appropriate response. Perceptual overload is often modulated by a perceptual attention mechanism that filters and prioritizes incoming data. Most existing attention mechanisms are tailored to the specific task the system is performing. A general-purpose attention mechanism must have a task-independent interface for controlling attention; support a heterogeneous set of sensors; support heterogeneous methods for processing sensor data; and support real-time throughput constraints. The CPA is a general-purpose attention mechanism that supports multimodal perceptual attention. Using it, an intelligent system can enact and control a variety of attention policies for any type or combination of sensor or sensor data. An intelligent system dynamically creates multiple heterogeneous perception tasks in accord with behavioral goals and installs them in the CPA. The CPA supports two general categories of perception tasks: detectors, which do not retain information between perception cycles; and trackers, which do. Perception tasks are prioritized using an attention policy and are executed using a priority-based scheduler. A wide range of attention policies can be expressed in this mechanism, including policies that dynamically modify perception priorities, policies in which emergency input overrides normal perception processing, and policies that dynamically change the level of resistance to perceptual distractions. Results show that perception intervals as short as 100 milliseconds can be achieved with a five-sensor robot under a variety of attention policies . Analysis of the system's performance under perceptual load shows that qualitatively different attention policies can be realized in the attention mechanism. We show that intelligent systems can use the CPA to implement the four primary characteristics of human perceptual attention: selective attention, sustained attention, divided attention, and top-down control.
PhD Thesis, Department of Computer Science, University of Texas at Austin.