Brains as Models of Intelligence

by Eric Hart

Intelligence testing has, for some time been in disrepute. Critics have a number of complaints with it; it is "culturally biased", it "fails to account for creativity", there are "too many kinds of intelligence to measure on one kind of test", and so on. But advances in the theory of computation, by enabling a general mathematical description of the processes which occur within human brains, indicate that such criticisms may overlook more general parameters of intellectual ability... i.e., that there exist mathematical models of human mentation which allow at least a partial quantification of these parameters.

Neural networks are computative structures analogous in general topology and functionability to the human cortex. The elements of such networks act like generalized brain cells linked by excitative and inhibitive "synapses" and conditioned by "learning functions" acting on "sensory" input. Because they are computative machines, they obey the same principles as machine programs, digital computers, and cellular automata.

While there are certain aspects of human brains which distinguish them from generalised neural nets, it is instructive to assume a close analogy between them; specifically, that everything a brain can do has its formal counterpart in the model. Thus, an algorithmic description of mental processes, translated into the "language" of the model, generates patterns with definite consequences in the model's descriptive logic.

On one level, the mathematics is identical with that of vector spaces and obeys the principles of matrix algebra; on another, it is statistical. But just as vector algebra and statistics have deep connections with other branches of mathematics, the conventional analysis of neural nets can be elucidated by the application of extended formalisms. In light of certain trivial insights concerning the model - e.g., that it is capable of localized self-simulation - this can lead to some enlightening conclusions on the structural and dynamical requirements of cerebral efficiency.

In principle, it is possible to measure deterministic intellectual potential by neurological analysis, provided we know precisely what to look at (some have tried, along these lines, to narrow it down to certain sets of mental reflexes, but things are not that simple). Of course, it is entirely possible that human intelligence has a nondeterministic component, as would be the case if humans possess anything remotely resemblihg "free will". But the potential for ratiocination should be measurable by means approaching anatomical description in objectivity.

It has long been a platitude that human beings use only a fraction of their brains at any given time. But it may be less important how many cells are firing than the patterns in which they fire; and the organization and density of neural populations are obviously crucial to the complexity and variety of procedural topologies, we are thus unlikely to find that everyone has been dealt the same intellectual hand in life, and this admits of intellectual distinctions among people. Eventually, a biochemical extension of the classical model should allow understanding and control of the roles that factors like emotion, motivation, and neural plasticity play in our mental growth and activities.

Because the commonly recognizable components of intelligence, such as deduction, association, and analogy, are all formalizable in terms of the model, its investigation bears strongly on the quest for true artificial intelliqence. This strengthens the hope that we may soon design machines which can enrich our lives and enable us to handle the ultracomplex systems which lie beyond our current abilities...even as we extend those abilities to achieve a better understanding of the universe within and without us. In a sense, many of our fondest dreams live in that one hope. Intelligent biological, ecological, and socio-economic simulators may one day construct algorithms for the indefinite maintenance of youth and good health, restore and stabilize the planet's frail ecosystems, and help us eradicate a multitude of societal ills. Many aspects of their structure and programming are already known. Whether these intelligent simulators are human, artificial, or bio-artificial hybrids, the study of intelligence is essential to their emergence and coexistence with us.

Due to the ultimate equivalence of all deterministic machines, my own study of neural nets has clarified aspects of cellular and conventional automata as well as deterministic systems in general. It also promises certain refinements in the usual methods intelligence testing.

The Mega Society

Copyright © 1989 by the Mega Society. All rights reserved. Copyright for each individual contribution is retained by the author unless otherwise indicated.