Mike Mertsock

• March 2001
• High school senior paper

With information the most important commodity in the modern world, there is a growing demand for complex and adaptable solutions to challenging problems. Artificial intelligence (AI), in simple terms the machine equivalent of human intelligence, offers the ultimate answer to many problems of today’s society. Though limited, specialized intelligent machines and software exist today, in broad terms none approaches the intelligence of a human. Achieving such intelligence in machines will unlock the awesome potential of the marriage of man and machine.

Current attempts at true AI have not been successful. However, promising new methods are developing that take advantage of the self-emergent complexity of biological systems, including self-directing networks of cell-like units, and genetic and evolution-based systems. The solution to AI may lie in these complex “emergent” systems, by emulating the structure and development of natural intelligence more directly than engineering more deterministic systems common to most current software. However, if these emergent systems succeed in creating AI, we may not fully understand their processes, a projection only reinforced by our slow progress in understanding the human brain.

Artificial Intelligence theory began early in the development of modern computers. The now familiar term first appeared at MIT in 1956 (Introduction 1). AI attempts to simulate the complexity and adaptability of the human brain. Though there are many who assert that the human brain is fundamentally no different from a computer and is therefore possible to simulate on a machine, computer scientists have encountered many difficulties in trying to create the digital brain. One such difficulty is the sheer complexity of the human brain’s connections: it has about one trillion neurons, each with an average of ten thousand connections to other neurons (Hillis 138). More importantly, human intelligence and behavior differs greatly from that of any current computer. Computers are more or less told exactly how to act by programmers and cannot adapt to situations other than those for which they are programmed. However, computers far surpass humans in mathematical computation.

AI research began around 1950, when British scientist Alan Turing proposed the now-famous Turing Test. He proposed that a computer could be described as intelligent if, in a question-and-answer session, an interrogator cannot distinguish between the dialog of the human and of the computer. After fifty years of research, there has been some progress, but an intelligent machine still does not exist.

The difficulty in creating AI is apparently not in the processing power of computers. Current computers are hundreds or thousands of times more powerful than the first digital computers, yet there is no real difference in the tasks they can do; newer computers are simply faster at accomplishing the same limited tasks. Many of the real steps towards AI were taken early on: for example, voice recognition is only now becoming widespread, yet the first such program was created in 1952 by Bell Labs (Masci 997). Even with the amazing speed of today’s computers, the best level of intelligence we can currently create is about equal to that of an insect. Apparently, there is something fundamental lacking in our approach.

There has been much current research in biology and computer science focusing on the nature of intelligence and finding exactly how biological intelligence works. Many findings suggest that there is a basic difference in the organization and operation of natural “computers” and man-made computers. No reasonably complex organism has a brain that was engineered, neuron-by-neuron, into a specific framework. Rather, processes such as learning and natural selection develop connections over time. To an extent, the brain is a self-organizing system. Almost all current computer programs, on the other hand, were designed line by line into a very specific, deterministic system. This severely limits a computer’s ability to adapt and to learn new abilities. Another important difference is in structure. A human brain is in one way very complex with its trillions of neurons and a confusing mass of connections in every direction. Yet, it is also quite simple. All the connections are more or less made locally by individual neurons. Thus, completely different areas of the brain can have almost indistinguishable patterns. Hillis states, “If there is a systematic difference between the wiring patterns in the brain’s two hemispheres, it is too subtle for us to discern.” (140) A brain is not systematic in the way that a computer is; a computer program follows a well-defined, linear path. A brain can be said to be nonlinear: it not only follows many indirect paths, these paths are usually hard to define.

These differences may seem disappointing, but a number of programming techniques have appeared lately which promise to achieve the complexity and nonlinear nature of the brain. The strength of each of these is that they directly simulate biological processes.

Neural networks are being used in pattern-recognition applications such as speech recognition. They directly simulate the process of creating and modifying connections between the neurons of a brain, creating their own rules as they gain experience. The solution is created by the interaction and collaboration of many individual neurons. Because the results are not always perfect, these systems are often “trained” by a human, who reinforces the behavior generating good solutions, as in the learning process. Over time, the system becomes more accurate and efficient as connections are modified.

Another technique simulates cells or simple animals rather than individual neurons. In such a system, the individual units are often identical and follow a simple set of rules. A group behavior emerges, and cooperation between the units creates an adaptable problem-solving system. Many applications of this method are being tested physically. Ant-like robots behave as a single colony to gather food more successfully than single ants. Other applications include the development of “smart” materials made of many simple, very small electronic components. A MIT study on such processes asserts that “we can learn much from living things’ ability to dynamically organize arrays of initially identical cells into highly ordered arrays of differentiated organs and to interconnect these systems in organized ways”. (Abelson 1)

A third promising approach is genetic algorithms, which “mimic within the computer the process of biological evolution” (Hillis 145). A programmer creates an initial set of similar programs and then applies principles of evolution to weed out the most successful programs. After several stages of random mutation, replication, and combination of these programs, a better group of programs emerges. In an example of the power of this method, Hillis created, using genetic algorithms, a sorting program that was faster than any program a human could write, and in an unexpected twist, he could not understand how it worked. These two seemingly disparate results could be a double-edged sword. While genetic algorithms promise to create more intelligent programs, Hillis remarks, “If evolution can produce something as simple as a sorting program which is fundamentally incomprehensible — it does not bode well for our prospects of ever understanding the human brain.” (147) But we may not have to understand the brain to be able to recreate it.

The success and potential of these techniques lie in the fact that, like their biological counterparts, complex behavior emerges from some combination of simpler entities. We can achieve the overwhelming complexity apparently inherent in intelligence by having the computers create the complexity themselves. However, there is still much work to be done:

The missing link, say many scientists, is that we neither understand nor place enough value on the models of computing that are all around us — and for that manner, inside us. These biological models are more complicated than the ones employed in today’s computing systems, to be sure, but also far more powerful and flexible. (O’Malley 62)

There have been many tangible results in applied programming from these techniques, but their full potential is not yet realized; most have only existed for only a few years. The possibility of AI resulting from one of or a combination of these technologies seems good. Even now, our society is immersed in the products of AI research. Thermostats, brake systems on cars, robotics, computer software, customer service departments, and the Internet all use applications of AI research to make our lives easier.