274 CATALYZING INQUIRY
8.3.3.1 Neuroscience and Architecture in Broad Strokes,
The most general lesson is that much of human cognition depends on the ability to ignore most of
the information made available by the senses.^74 That is, a very high fraction of the raw information that
is accessible through sight, sound, and so on does not participate directly in the human’s cognitive
processes. Human and mammalian cognition is based on an architecture that involves a flexible, but
low-capacity, working memory and attentional selection mechanisms that place events and objects into
working memory where they become available for cognitive processing.^75
This approach of selective attention stands in sharp contrast to traditional algorithms that are
designed with the goal of seeking optimal solutions and based on the use of as much information about
the problem domain as possible. The architecture of biological computation has generally evolved with
a different purpose—the adequate management of a complex, changing, and potentially dangerous
environment in real time (where “adequate” means “provides for survival”).
This architecture is based on a two-track processing arrangement—a very flexible, albeit slow
system that implements consciousness, awareness, and cognition but attends to only few things, and a
large number of online, fast-acting, sensory-motor systems that bypass attention and awareness (e.g.,
eye movements, head and hand movements, posture adjustments, and other reflex and reflex-like
responses).
Koch et al. have investigated the utility of such a strategy in multiple contexts: (1) a saliency-based
visual attention mechanism that selects highly “salient” location in natural images for further process-
ing;^76 (2) a competitive, two-person video game in which an algorithm that focuses on a restricted
portion of the playing field outperforms an “optimal” player when a temporal limitation is imposed on
the duration of each move;^77 and (3) an algorithm that rapidly solves the NP-complete bin-packing
problem under most conditions.^78
8.3.3.2 Neural Networks,
Biology affords an alternative computing model that (1) appears well suited for many ill-posed
problems constrained by uncertainty, which is the problem set for which digital machines to date have
been reasonably ineffective; and (2) provides an existence proof that slow and noisy circuits can under-
take very rapid computations of a certain class. Furthermore, it provides huge numbers of working
examples. Although the mechanisms underlying nerve tissue computation are not well understood
despite many decades of study, the fact remains that biology has found incredibly good solutions to
many engineering problems, and these approaches may well serve to inform practical solutions for
engineering problems posed by human beings. Indeed, although biological tissue is not naturally suited
for information processing as understood in traditional terms, the fact that biological tissue can do
information processing suggests that the underlying architectural principles must be powerful indeed.
Neural networks are among the most successful of biology-inspired computational systems and are
modeled on the massively parallel architecture of the brain—and on the brain’s inherent ability to learn
(^74) C. Koch, “What Can Neurobiology Teach Computer Engineers?” January 31, 2001, unpublished paper, available at http://
www7.nationalacademies.org/compbio_wrkshps/Christof_Koch_Position_Paper.doc.
(^75) F. Crick and C. Koch, “Consciousness and Neuroscience,” Cerebral Cortex 8(2):97-107, 1998.
(^76) F. Crick and C. Koch, “Consciousness and Neuroscience,” Cerebral Cortex 8(2):97-107, 1998; L. Itti and C. Koch, “A Saliency-
based Search Mechanism for Overt and Covert Shifts of Visual Attention,” Vision Research 40(10-12):1489-1506, 2000; L. Itti and C.
Koch, “Target Detection Using Saliency-based Attention,” Search and Target Acquisition, RTO Meeting Proceedings 45, NATO,
RTO-MP-45, 2000; L. Itti, C. Koch, and E. Niebur, “A Model of Saliency-based Visual Attention for Rapid Scene Analysis,” IEEE
Transactions on Pattern Analysis and Machine Intelligence (PAMI) 20:1254-1259, 1998.
(^77) J.G. Billock, “Attentional Control of Complex Systems,” Ph.D. Thesis, 2001, available at http://sunoptics.caltech.edu/~billgr/
thesis/thesiscolor.pdf.
(^78) J.G. Billock, D. Psaltis, and C. Koch, “The Match Fit Algorithm: A Testbed for the Computational Motivation of Attention,”
International Conference on Computational Science 2: 208-216, 2001.