Selective Attention as an Optimal Computational Strategy
Greg Billock, Christof Koch, and Demetri Psaltis ABSTRACT
We explore selective attention as a key conceptual inspiration from neurobiology that can motivate the design of information processing systems. In our framework, an attentional window, the “spotlight of attention,” contains some reduced set of data from theenvironment, which is then made available to higherorder processes for planning, real-time responses, and learning. This architecture is invaluable for systems with limited computational resources. Our test bed for these ideas is the control of an articulated arm. We implemented a system that learns while behaving, guided by the attention-based content of what the higher-order logic is currently engagedin. In the early stages of learning, the higher-order computational centers are involved in every aspect of the arm’s motion. The attentionally assisted learning gradually assumes responsibility for the arm’s behavior at various levels (motor control, gestures, spatial, logical), freeing the resource-limited higher-order centers to spend more time problem solving. remarkable fact—documentedthroughout the book— that only a very small fraction of the incoming sensory information is accessible, in a conscious or unconscious manner, to inﬂuence behavior. Many people have speculated about consciousness and its function. According to Crick and Koch (1988; Koch, 2004), the function of conscious visual awareness in biological systems is to “[p]roduce the best current interpretation of thevisual scene in the light of past experience, either of ourselves or of our ancestors (embodied in our genes), and to make this interpretation directly available, for a sufﬁcient time, to the parts of the brain that contemplate and plan voluntary motor output, of one sort or another, including speech.” This representation consists of a reductive transformation of the massive, real-time sensory inputdata. That is, the content of awareness corresponds to the state of cache memory that holds a compact version of relevant sensory data as well as recalled items. This strategy can deal with more complex scenarios and generate a strategy for action (Newman et al., 1997). This ﬂexible, but slow, aspect of the system, is complemented by a set of very rapid and highly specialized sensorimotor modules(D. Psaltis, personal communication, 1995), “zombie agents” (Koch, 2004), that perform highly stereotyped actions (e.g., driving a car, moving the eyes, walking, running, grasping objects). Figure 4.1 illustrates one way in which these cognitive strategies may be mapped onto a machine architecture (Billock, 2001). The sections of the diagram toward the bottom—the motor/processing modules, earlyprocessing, and error generation—reside below the level of awareness, with fast reﬂexes and extensive procedural memories. Selective attention and aware-
I. THE ATTENTION–AWARENESS MODEL: AN INTRODUCTION
Computers and software have recently joined the long line of human tools inspired by biology. In this case, it is the phenomenal capabilities of biological nervous systems that intrigue andchallenge us. Our desire to mimic the brain stems from the abilities it possesses, which are in so many cases superior to those we can implement today. We here explore the extent to which attentional selection can convey functional advantages to digital machines. By attentional selection we refer to the
Neurobiology of Attention
Copyright 2005, Elsevier, Inc. All rights reserved.II. LEARNING MOTION WITH AN ARTICULATED ARM
Early Processing Processing Modules Environment
Error Signal Generation
FIGURE 4.1 In this functional model of the role of attention and awareness, the pathway incorporating the attentional bottleneck operates in parallel with the faster...
Leer documento completo
Regístrate para leer el documento completo.