honavar@SPEEDY.WISC.EDU (A Buggy AI Program) (11/24/87)
In article <4739@wisdom.BITNET> eitan%H@wiscvm.arpa (Eitan Shterenbaum) writes: > >a) You can't understand the laws under which a system works without > understanding the structure of the system ( I believe that our > intelligence is the result of our brain's structure ) Not entirely true. We can often gain insights into what structures are needed to produce a certain observed behavior simply by observing the system's behavior. This would then enable us to hypothesize about the structures that actually produce such behavior. We would then test the hypotheses by putting them through experimental validation. Just as one can have several different computers that are functionally equivalent, it is reasonable to expect that there several possible architectures (the human brain being one of them) that are capable of intelligence. > >It seems to me that > 1) You have no definition for Intelligence. > 2) You want to have the rules of Itelligence. > 3) Thus you build systems inorder to simulate Intelligence. > 4) Since you don't know you're looking for and since you have no > basic rules to simulate the intelligence on, you invent your > own local definition and rules for Intelligence. > 5) Then youtry to mach your results with your expectations of what > the results should be. This is an oversimplified view of the research methodology in AI and Cognitive sciences. It is true that we don't have a good definition of intelligence. For purposes of AI, it is sufficient to say that we want to build systems that exhibit the kinds of behavior that are believed to require intelligence if performed by humans (I forget the author that first suggested this definition of AI). This is an operational definition or at least a basis for an operational definition of artificial intelligence. Given this, there are several alternative approaches one could adopt in building intelligent systems - including the one of simulating a system that most of us agree is capable of intelligence, the human brain (plus the sensory mechanisms). The search for architectures for intelligence is by no means an unconstrained, blind search. The hypothesis can be constrained by utilizing data gathered from experimental research in psychology, neuroscience, and related areas as well as theoretical analysis of complexity of the tasks involved and so on. > >Correct me if I'm wrong but I do feel that the neuro-biologists chaps are >in the right track and that the Computer scientists should combine efforts >with them instead of messing around with AI. > I agree that AI researches can benefit from the research findings in neuroscience. It is also true that computational theories advanced in AI can provide insights to neuroscientists as well. In fact, there is evidence of this interaction in the works of David Marr, Shimon Ullman, and others. Cognitive psychology is another field which is at least as relevent as neuroscience to work in AI.