Toward human level machine intelligence - is it achievable?

Lotfi A. Zadeh
2008 2008 7th IEEE International Conference on Cognitive Informatics  
Science deals not with reality but with models of reality. In large measure, scientific progress is driven by a quest for better models of reality. The real world is pervaded with various forms of imprecision and uncertainty. To construct better models of reality it is essential to develop a better understanding of how to deal with imprecision and uncertainty. Such understanding is a prerequisite to achievement of human level machine intelligence. 3 /89 LAZ 4/7/2008 PREAMBLE As an issue, human
more » ... evel machine intelligence, or HLMI for short, has a special significance for me. When I started my teaching career at Columbia University in 1950, I wrote an article entitled "Thinking machines-a new field in electrical engineering," published in Columbia Engineering Quarterly, January 1950. The article was written seven years before the birth of AI in 1956. I have been an observer of progression toward HLMI ever since. 4 /89 LAZ 4/7/2008 CONTINUED The 50's of last century were the years of unbounded enthusiasm for AI and exaggerated expectations. In my article, I referred to an article headlined "Electrical brain capable of translating foreign languages is being built," which was published in the late 40's. Today we have translation software but nothing that can compare in quality with human translation. 5 /89 LAZ 4/7/2008 CONTINUED In 1948, on the occasion of inauguration of IBM's Mark1 relay computer, Howard Aiken, the Director of Harvard's Computation Laboratory said, "There is no problem in applied mathematics which this computer cannot solve." What could be more unrealistic? However, we should bare in mind what Jules Verne, the famed science fiction writer, wrote at the turn of the 20 th century: Scientific progress is driven by exaggerated expectation. 6 /89 LAZ 4/7/2008 CONTINUED Where do we stand today? AI can point with pride to important successes but progress in the realm of human level machine intelligence has been limited and slow. Anyone who had the painful experience of struggling with a dumb automated customer service system will readily agree that HLMI is not yet a reality. The Turing test lies far beyond. Today, no machine can pass the Turing test and none is likely to do so in the foreseeable future. 7 /89 LAZ 4/7/2008 CONTINUED What is the problem? What is holding up progress toward achievement of HLMI? In an article entitled "A new direction in AItoward a computational theory of perceptions," AI Magazine, Vol. 22, No. 1, 73-84, 2001, I pointed out that humans have two remarkable capabilities. First, the capability to converse, summarize, reason and make rational decisions in an environment of imprecision, uncertainty, incompleteness of information, partiality of truth and partiality of possibility. 8 /89 LAZ 4/7/2008 CONTINUED And second, the capability to perform a wide variety of physical and mental tasks, such as driving a car in city traffic, without any measurements and any computations. Underlying these capabilities is the human capability to reason with and act on perceptions rather than measurements. A natural language is basically a system for describing perceptions. Mechanization of this capability is a prerequisite to achievement of human level machine intelligence. 9 /89 LAZ 4/7/2008 CONTINUED In my view, the lack of machinery for dealing with perceptions is the prime reason why progress toward achievement of HLMI has been limited and slow. What is needed to develop a computational theory of perceptions? In the past, AI's armamentarium consisted, in the main, of methods based on classical, Aristotelian, bivalent logic. The problem-solving capability of AI was significantly enhanced through the addition of probability theory to its armamentarium. 10 /89 LAZ 4/7/2008 CONTINUED But the problem is that the development of a computational theory of perceptions is beyond the reach of bivalent logic and bivalent-logic-based probability theory. I have been arguing for some time that what AI has to do is to add to its armamentarium concepts and techniques drawn from fuzzy logic. I believe that sooner or later this view will find acceptance because the inadequacy of existent tools will become too apparent to ignore. 11 /89 LAZ 4/7/2008 CONTINUED My AAAI article contains a key idea. Specifically, the idea is to compute not with perceptions per se but with their descriptions in a natural language. In this way, computation with perceptions reduces to computation with information described in natural language, NL-Computation or equivalently, Computing with Words (CW) for short. 12 /89 LAZ 4/7/2008 CONTINUED A prerequisite to computation with information described in natural language is mechanization of natural language understanding. In turn, a prerequisite to mechanization of natural language understanding is precisiation of meaning 13 /89 LAZ 4/7/2008 CONTINUED What has been widely unrecognized is that in the final analysis, progress toward achievement of human level machine intelligence requires a resolution of a critical problem-the problem of precisiation of meaning. To resolve this problem what are needed are new concepts and new techniques. An outline of what is needed is presented in the following. 14 /89 LAZ 4/7/2008
doi:10.1109/coginf.2008.4639144 dblp:conf/IEEEicci/Zadeh08 fatcat:q6s7gaj3yvb6hl365wlhdvo7au