Publications

(This essay appeared in DM-list as a reply to R. Dybowski Call for Papers for ML Journal Issue on Fusion of Knowledge with Data, see item 31. GPS)

Dear Mr. Dybowski:

When I asked the DM list for a definition of data mining, I got enough interesting answers to scribble out several pages of notes. At the first level this exercise gave a very good answer to my question. At the second order of analysis, it tells us something about the problem of arriving at definitions in general. Definitions are ubiquitous in the testing of (real) human intelligence. Perhaps this exercise has something to say about your special issue as well. I assume that those giving the definitions were conscious of their answers as they typed them or verbalized them to a voice recognition program. I also assume that these definitions are not carried about in the conscious mind at all times. Those answering had to reach into the unconscious and retrieve the definitions or develop them. Perhaps then we should look more closely at the history of work in introspective psychology and the study of the unconscious mind.

I think the short essay below might be suitable as a letter to the Machine Learning Journal. If so perhaps a title would be "Knowledge Engineering: What happens when the human unconscious is made conscious and installed in learning machines?"

----------------------------------------------------------------------------

The 'call for papers' of this special issue of the Machine Learning Journal notes that "knowledge engineering and machine learning remain largely separate disciplines." How can they be "combined to construct decision support systems?" it asks. I am going to give one answer to that question by going back to nineteenth century introspectionism and the quest for a breakthrough in (real) human intelligence. I think this will provide some fruitful ideas on how we can advance current (artificial) machine intelligence and learning.

Chapter 13, "Psychology Becomes Self-Conscious" from Brett's "History of Psychology" (1912; 1921) tells us that "In German psychology in the nineteenth century two main trends were discernible which are of methodological interest. Psychologists were very preoccupied with the self or soul as a possible object of study; they were also very concerned about themselves and the status and terms of reference of their developing inquiries." (p. 533). If human subjects could look within themselves and tell programmers how their intellect works, great advances in AI and machine learning would be made. This would constitute knowledge engineering applied to everyday intellectual skill. Isn't this a reasonable quest for a species which goes by the name of "homo sapiens"? Further to that, what will the status of "robo sapiens" be when this undertaking is successful?

The self is largely unconscious as any layman will tell you today. Brett credits Von Hartmann (1842-1906) with developing the Doctrine of the Unconscious, particularly with his 1901 book, Die Moderne Psychologie, qualifying this by saying "Though to some of his contemporaries Hartmann appeared to be wholly original, two-thirds of the Doctrine of the Unconscious was already commonplace." (p. 578). Brett's history speaks highly of Fechner. "The study of Fechner's life is one of the most instructive ways of following the progress of thought in the nineteenth century...in the year 1860, the Elemente der Psychophysik was completed." (pp. 580-581). Given that "His Psychophysik was not the mere invention of a science; it was the attainment of a new plane of thought" (p. 584), we need to take it seriously. Though every psychology student since that era has had to memorize the Weber-Fechner Law, my interest presently has to do with the general methodology of introspectionism and not its nineteenth century findings. Most succintly this methodology is expressed in the following: "The last source of knowledge is Einfuhlung or self-objectification." (p. 609).

The problems of machine learning and AI today are still largely the problems of self-objectification and the turning of unconsciousness into consciousness. For example, how can we develop software to clearly articulate (a) reading skill; (b) everyday conversational skill; (c) object recognition skill? All three are everyday skills or abilities. We take them for granted. Yet they are largely unconscious to us. To appreciate how remarkable our agnosia is, consider other skills by contrast. Imagine a bus driver who says he does not know how he drives the bus...it 'just happens'. Consider your surprise upon hearing your dentist profess to have little conscious knowledge of how dentistry is done. Isn't it amazing then that everyday skills like those of reading (with comprehension), conversing and recognizing objects are mostly of an unconscious nature when it comes to HOW we do them? Homo sapiens indeed.

The reward for an understanding of "The Psychology of Everyday Things" (the title of a book by computer scientist/psychologist Donald Norman) would be be no less than a comprehensive understanding of how to write software for intelligent machines which will meet or exceed human equivalency. In this case I mean reading, conversing and the recognizing of objects as 'everyday things'. Individual Differences, as I note in my 1976 text by that name (with A. Buss; Gardner Press) are ubiquitous. Every clinical psychologist knows too that some patients are more insightful than others (eg, see Poley, Lee and Vibe, "Alcoholism: A Treatment Manual", 1979, Gardner Press). It is therefore my expectation that if a notice were posted on the world wide web seeking assistance from those who can consciously articulate that which is unconscious to most people, most of the time, the effort would produce good results. Whether by natural talent or training or both, those with exceptional introspective abilities with respect to "everyday things" are likely to be in the world population. The solution to the presenting problem of knowledge engineering and machine learning then is suggested via social psychology and marketing psychology as much as anything else. To find and engage the assistance of those who can do knowledge engineering in these domains would have to be posted by a well known and authoritative body to be effective, eg ACM (Association of Computing Machinery).

Skinner's dictum "If it can be verbalized, it can be programmed" is true or mostly true. I have yet to see a definitive answer as to whether he meant this to include both the broad and narrow (computing science) definition of "programmed" but I will assume that the answer is yes. After all, the Mark I computer was built at Harvard in the 1940's and Skinner must have understood what computing science meant by that word. Lets find those gifted people who can verbalize for us how they read with comprehension, converse and recognize objects. When they give us the benefit of their knowledge engineering in the domain of everyday intellectual skills, we will have some valuable software. The software will be verbalized or written by them in EL, "Everyday Language". Since a standard computer language like LISP is just chicken scratchings without meaning, I think we can call a script of meanings in EL, "software" as much as the LISP or Fortran or COBOL into which it is transcribed. Thus we would be looking for software writers in the world population who have never thought of themselves in these terms. Some marketing advantages accompany this.

Once this is successful, there will be a "fusion" of domain knowledge (in the domains of reading, conversing and recognizing objects) with the data to which the domain knowledge is applied. Add that capability to existing capabilities in memory and logic/arithmetic/mathematical ability and you have powerful software for "decision support". It will be so powerful that it should meet or exceed the Moravec criterion of human equivalency for the achieved products of factors of primary mental ability in humans (see Ch.3 of "Individual Differences"). In other words we will have a machine with "superhuman AI". Even if the introspectionists in this project are limited, they will achieve partial results and those partial results may enable less gifted people to discover some rules for further knowledge engineering.

In closing, let's describe the successful outcome of such a project in terms which are as dramatic as they are realistic upon successful completion. Imagine humanoids like the recent Honda and Sony prototypes programmed with such capabilities. When I coined the expression "robo sapiens" I had in mind the prospect of human equivalency in intellectual ability. But a human look-alike dramatizes the concept, especially if the humanoids can walk among humans and are given at least a degree of autonomy. Imagine ASIMO being able to carry on an everyday conversation as well as a typical human, answer questions about any text as well as a human and able to recognize objects as well as a human. In my opinion, gifted introspectionist humans would enable us to make significant progress in that direction. And few doubt that robo sapiens will attain human equivalency some time in this century. Thus the prospect of trying to attain it with a mega-project planned now for completion over the next decade and with a budget comparable to that of the International Space Station should not be ruled out. When homo sapiens becomes fully conscious and programs robo sapiens with that consciousness, what will the theologians and philosophers say? Is there any reason robo sapiens, particularly after a few generations of learning and self-improvement will not confound us with the answers? What will we say when robo sapiens tells us it is conscious and gives better lectures and answers on what it means to be conscious than any human?