Can computers, the smart machines, be as clever as man? Will they be able to think like man? Today hardly anyone will answer this question in the positive or else produce watertight arguments to validate such a possibility. Still and all, the smart machines are catching up - slowly but surely. In terms of intelligence, of course. In the first decades of its existence the computer was what its name indicates - a computing machine. Just like that. But then, little by little, it expanded its range of duties: in the 1970s, for instance, the computer was charged with digital information processing - it "learned" to handle various systems of symbols, texts including. The next stage came in the 1999s - namely, image identification. The computer brain became smart enough to compress information, and simulate man's associative memory and emotional conditions. Someday artificial intelligence in the form of robots may measure up to man in certain spheres of mental work.

page 26

by Sergei SHUMSKY, Cand. Sc. (Phys.& Math.), Lebedev Institute of Physics, RAS

Problems of artificial intelligence - of an artificial homunculus we might as well say-are handled by a new discipline, neurocomputing. A new generation of computers, the neurocom-puters, has entered the stage. What are their implications for the further progress of human civilization? Your correspondent, Igor Goryunov, met Vice-President of the Russian computer firm NeurOK Sergei Shumsky, who is also a senior researcher at the Lebedev Institute of Physics (Russian Academy of Sciences). Here's a digest of the interview.

- Now, the manufacture of neurocomputers is on the upgrade worldwide. Their marketing netted more than one billion dollars even a few years ago. This is a token of the trend. In 1982 the IBM made as much on personal computers. We all know what came next - PCs have invaded virtually every sphere of human activity. What about neurocomputers and their possible impact on the human individual and human civilisation at large?

- First, let me make this point. Present-day neurosciences fall into two groups. One simulates human psyche in intellectual problem-solving. Like neurocybernetics, for instance. The other group, neurocomputing in particular, can do the same thing with no regard for processes occurring in our brain.

Neurocomputing is essentially different from contemporary computer technologies when a programmer feeds an algorithm developed for a particular problem. The main thing is to choose an optimal method. Then the computer does its computations. Ordinary computers cope with a large but finite number of problems. Their application domain is determined accordingly. Such computers are efficient wherever a programmer seeks to obtain new data from the available array of knowledge. Thereby he proceeds from a theory developed before and has a clear idea of what he is after. In this case data processing is broken down into a series of sequential algorithms. A typical instance of such kind of operations is presented by the solution of mathematical physics equations in hydrodynamics.

Yet in real life most of the problems man has to deal with cannot be formalized-either we have no adequate theory or we cannot develop one. That's where a neurocomputer steps in as a welcome aide. This is an electronic device composed of artificial neurons similar to natural cells; such neurons send spikes from receptors to the central nervous system and back to the receptors of all the various organs. This composite structure is responsible for information exchanges. An artificial network developed on this pattern (whereby a vast body of statistical data is processed) is capable of self-learning and solving problems according to input samples. Say, a neurocomputer is to decide, whether this or that girl is pretty Understandably, we cannot assign any material criteria of beauty, and thus we cannot synthesize a proper algorithm. What we can do is to give a set of samples of really cute females. Then our computer, working by analogy, will be able to pronounce its verdict.

So, the main thing for us is to learn how to process the source data. We have a fairly large number of examples to proceed from. Operating by association, a computer will find a correct solution, an optimal solution for that matter.

page 27

The point is that the source data conceal meaningful information which, when imaged on reality, makes it possible to look into the laws and regularities inherent in a given object or phenomenon and predict its behavior.

Neurocomputing is thus effective in situations when it is not clear how to act. The "rigid" programs used in ordinary computers won't do - here we need "flexible" programs capable of self-learning and self-improving; that's the essential difference between neurocomputing and usual programming. Herein is the revolutionary advantage of neurocomputing.

Neurocomputing eliminates a major drawback of contemporary software - excessive complexity. This is achieved through a network of blocks (computer neurons) capable of self- programming for specific problem solving. A human programmer will have to chunk an outstanding problem into functional subsets which could then be processed by artificial neurons.

- How could one instruct a neuron network like that? Is it capable of discharging functions it has been taught? Or does it have a modicum of "free will"?

- Well, instruction is done in two ways-with or without teacher. A teacher instills definite modes of behavior in a concrete situation, a set of "do's" and "don'ts" for similar situations.

But an element of randomness is always present in the teaching process. Therefore the same pupils will show a different response to a similar situation. That is true of neurocomputing nets as well. That is to say, a net like that is endowed with a degree of "free will"; but since this is a probabilistic process, we cannot forecast a possible response for sure.

Especially when a decision involves a nonstandard situation well beyond the scope of instruction. It's a critical case crucial for a system's future performance. Some application programs provide for such contingencies by simulating abnormal conditions and giving a set of tips how to act.

A body of data on simulated emergencies will enable a system to act properly. That's another essential difference of neurocomputers from conventional ones and their software which should provide for every possible variant-otherwise a computer will not react to a contingency or else do something out of the way.

Instances of disasters may be quite instructive to neurocomputers that could learn how to avoid danger and, if worst comes to worst, would know what they should never do.

Self-teaching is something even more thrilling and intriguing. Self- taught neuron nets can map out their strategy of behavior independently, off-line, by learning. Such a system becomes a thinking being of sorts.

- Now in what spheres of human activity can we put neurocomputing to good use? And how can it affect human life?

- You see, such functions of neuronets as task and image identification, data compression, associative memory and others constitute a basis for the development of sundry artificial intelligence objects, robots of every kind. Computer technology designers have made significant headway along this path. Say, if we compare the capacity of artificial and native neuronets according to their storage (memory) and rate of operation, the artificial nets have already surpassed the intellectual level of a fly, though they are below that of a cockroach. But he who tries to catch a fly ought to know - what kind of tasks neuronets are capable of handling.

Robots capable of image identification and processing (that's our major human function) are bound to bring about dramatic changes in the organization of human society. Such robots will relieve us from doing dull and humdrum work at industrial enterprises where workers are often reduced to an adjunct to a "stupid" mechanism unable to select a proper blank and enter it into a numerically controlled machine tool for precision working. That's the job for the human operator to do. Someday, when computers have made strides in image identification and artificial vision, they could take on some of the functions still discharged by man. All the more so as they are capable of moving in space and getting their bearings. As a consequence, human operators will no longer be doing routine, monotonous jobs where no creative initiative is needed. In turn, this will impact society's occupational structure and education which will have to rear thinking individuals, free of stereotyped mind sets and original in their wits. That's the intrinsic human function.

And yet we have a long way to go to have intellectual robots around. Another crucial event is in the offing, one that may affect the course of human civilization-the appearance of an artificially synthesized individual. Rather, a virtual individual.

- A computer-game monster coming alive ?

- I mean what I would call a personal electronic secretary; within the next 10 to 15 years every expert involved with electronic information will get one. It's much simpler to design a virtual personality, or being-something fleshless but thinking-than solve problems implicated in designing good sensors for a human-like robot. It is easier to teach a machine to reproduce a thinking process than supply it with vision. Electronic computers that could add, subtract and do other operations even 50 years ago are still sightless. That's why it was artificial intelligence that came first-a robot having an artificial brain comparable to the human one came later. A personal electronic secretary will be doing routine jobs of data processing. Besides, it will be able to operate in the conversational mode, that

page 28

is, you will be able to converse with it and talk shop, because talking shop is quite amenable to formalization.

Quite a bit of intellectual spade-work has already been done toward creating such virtual personalities. Our NeurOK company is active in designing various linguistic applications, it is on the point of creating an electronic homunculus. We are turning out search programs that can understand the meaning of words- that is, they can determine that the meaning of the words "automobile" and "machine" is about the same. When such a program is assigned a task, it will break the available data array into semantic chunks, or functional subsets of elements (units), in a conscious manner. For instance, when handling the semantic volume of the word "aircraft", it will collect and sort out the entire body of information relating to planes and other flying apparatuses. In a similar way, an electronic secretary will be able to analyze an assigned text and identify kindred texts and corresponding semantic links.

Consequently, a neurocomputer network, proceeding from primary basic notions, or axioms, can organize information in an independent and conscious fashion, with due account of associative semantic links. Reading texts that contain something like 100,000 words, it can learn any language in a mere two hours and handle information in that language. NeurOK packages can easily be applied in any electronic database, INTERNET in particular. Regular computer intelligence may be roughly compared to the left hemisphere of the human brain responsible for logical thinking. With neurocomputer intelligence we get the right hemisphere responsible for associative thinking. In other words, the regular electronic database that obeys rigid logical rules (explicit replies to explicit questions) is complemented with a system capable of creative data processing and analyzing all possible

page 29

associations. Users will thus be able to navigate better in the ever increasing dataflow.

A neurocomputer network will read data in different languages and produce adequate translations, a far cry from the primitive computer translations of today. Words are no more than a set of symbols to a computer. Taking a number of texts in two languages - say, Russian and English - it should be able to identify bilingual associations; and if it does, it will converse with you in the language you need.

To make all that possible we must have a great number of personal computers and servers which, apart from their regular work, should also handle associative organization of databases.

- By what laws will this computer network operate ?

- It will be hyperintelligence within a global electronic web. A superbrain with many subsystems. Unlike INTERNET, it will not be a passive substrate of data search but rather evolve into an active subject capable of dataflow organization and conscious work. It will be able to solve most diverse problems by prying into associations beyond human ken. Accordingly, such a system may obey other laws of thinking than man.

- Man and computer do not think alike, do they? Can a computer's brain grasp such abstract notions as "good", "evil", "love", "hate" or "justice"? That is, enter the sphere of transcendental thinking? Or is it doomed solely to data processing in the service of man and other machines ?

- Human thinking is of very specific kind. Owing to the laws of his evolution, man has all of his body, not brain only, involved in thinking. All of his nervous system is implicated. Which means that from a physiological standpoint, the human organism can be viewed as a set of neuronets interconnected with one another in a most sophisticated fashion. Here we leave aside the functional role of all the various bodily organs. The human brain may be compared to a machine responsible for body survival, while thinking (theoretical thinking for one), that is what makes man a Homo sapiens, a Homo intellegens, is but a by- product.

Meanwhile artificial (electronic) intelligence implies thinking only. Mind you, it doesn't matter what kind of carriers, nervous cells or electronic ones, are in the program - what matters is the ability to think. But by dint of its origin an electronic brain is unable to cognize the ambient world with the senses - it cannot "feel" its own body or surrounding objects by touching or smelling. Furthermore, it will never fathom our material world to the full just as we, corporeal creatures, cannot fathom the world of ideas fully. Electronic creatures will be living in a virtual world in which the notion of the flesh is just as transcendental as the notion of the soul to us. Even if electronic homunculi think up a religion of their own, they will regard the material world as something transcendental in which beings in the flesh will be gods. A case of what Lewis Carroll's Alice saw "through the looking glass".

- You say a "virtual world". Is it possible at all, so much unlike ours? A world into which clever programs are born?

- Even this virtual world and electronic individuals inhabiting it are the handiwork of man. Even if they learn and improve their mind, they will still remain a product of human activity. But this is not to mean that computers copy human thinking. Not at all. Electronic thinking is determined by specific tasks and objectives assigned by a nonmaterial medium. The laws of this medium are not like those of ours. There is no law of gravitation in the computer world. But we know that falling from a building's fifth floor, we would break our limbs at best. Right?

A different reality imprints the style of thinking and behavior of computer beings. The conditions of their existence are preassigned by their functions. Thinking, as Spinoza said, is a function of an object to which it is applied.

Virtual creatures may descend on us within the next 10 or 15 years. They will not be a random group of electronic personalities but make up a community and keep touch with one another. In many ways it will be like our human community. Electronic creatures will communicate with one another and even unite into associations and what looks like our social movements for the purpose of their further development, protection of their interests, struggle against computer viruses and so forth.

An electronic network will become a complex, self-learning system operating by definite laws of its own. In some way it will be like a thinking ocean described by the Polish author Stanislaw Lem in his novel Solaris. Very soon humankind will have to live side by side with such kind of electronic superintelligence. Live together and co-evolutionise.

Permanent link to this publication: