Xref: utzoo sci.lang:1696 comp.ai:1163 Path: utzoo!mnetor!uunet!tektronix!sequent!mntgfx!msellers From: msellers@mntgfx.mentor.com (Mike Sellers) Newsgroups: sci.lang,comp.ai Subject: Re: the role of biological models in ai Message-ID: <1987Dec17.171519.3477@mntgfx.mentor.com> Date: 18 Dec 87 01:15:16 GMT References: <23@gollum.Columbia.NCR.COM> <2590@gryphon.CTS.COM> Organization: Mentor Graphics Corporation, Beaverton Oregon Lines: 78 Keywords: models, purpose of ai >In article <23@gollum.Columbia.NCR.COM> rolandi@gollum.UUCP () writes: >> >> According to some AI theorists, (see Schank, >>R.C., (1984) The Cognitive Computer. Reading, Mass.: Addison-Wesley) >>AI is "an investigation into human understanding through which we learn >>...about the complexities of our own intelligence." Thus, at least for >>some AI researchers, the automation of intelligent behavior is secondary >>to the expansion and formalization of our self-understanding. From what I've seen of AI research, this may not be true (in most cases). I think most AI researchers are not so concerned with self-understanding as they are with creating a program that interacts with humans in a seemingly intelligent way. It makes no difference if the methods or structures used bear any resemblance to the human way of doing things. I believe the problem for most active researchers is one of scale: you cannot possibly hope to create a program that models human cognitive processing, and you have to get *something* running, so you set your sights a little lower and brush aside questions of how well the program corresponds to humans. This is not meant to sound demeaning or even cynical, just realistic. >>This is >>assumed to be the result of creating computational "accounts" of (typically >>intellectual) behavior. Researchers write programs which display the >>performance characteristics of humans within some given domain. The >>efficacy of a program is a function of the similarity of its performance >>to the human performance after which it was modeled. Thus AI programs are >>(often) created in order to "explain" the processes that they model. The last three statements are, I believe, rarely (if ever, in "classical" AI research) true. In the vast majority of cases, we do not even know what the "performance characteristics of humans" are! For a task of any real complexity, modeling a human's performance (when it can be measured) is still a matter of theory and conjecture rather than programming (see the scale problem I mentioned above). For example, even for all their hype and worth, knowledge-based (expert) systems do not even begin to approximate the actions of a human expert. The most advanced projects in this area have some explanatory capabilities, and some skill at incorporating new or conflicting facts in their decision making process, but this is just scratching the surface of how human experts operate. Lastly, current AI programs are like the stork-story of human birth as far as explaining human behavior or cognitive processing goes; they may provide something that we can learn from later on, but they do not really get us any closer to knowing what is really going on. In article <2590@gryphon.CTS.COM>, sarima@gryphon.CTS.COM (Stan Friesen) writes: >My problem with this class of AI research is that I question it >validity/usefulness. Why should there be only *one* algorithm for a >particular 'behavior'? What evidence do we have that the algorithms that >we are writing into our programs are in fact related in any way th the >ones used by the human brain? Mere parallel behavior is NOT sufficient >evidence to claim increased understanding of a human behavior, some >evidence from neurology and psychology is necessary to at least >demonstrate applicibility. In particular, I find most current AI >algorithms to be far too analytical to be realistic models of human, >or even animal, cognition. Most AI algorithms have little if any resemblance to how humans function. How important this fact is depends on who you talk to. Of those people doing research in PDP (parallel distributed processing, or artificial neural networks, or connectionist nets, etc), many are convinced that some correspondence with the functioning of the human brain is important (possibly vital). This is not to say that this way of operating is the "only way". It is, however, the only way that we know of. Later, when we have all the principles behind cognition down pat, we can begin to branch out in different directions. Interestingly, many of the people doing this research are psychologists and neurologists, so there is (hopefully) an increasing amount of knowledge and techniques from these fields being used in this research. For the time being, however, the level of cognition we will be seeing arising from PDP research will be more reminisicent of a flatworm or a sea slug than a dog or a human (I predict, however, that this is more than we will see from more "classical" AI methods, which will continue to be more concerned with outward function than with inward correspondence). -- Mike Sellers ...!tektronix!sequent!mntgfx!msellers Mentor Graphics Corp. Electronic Packaging and Analysis Division