gilbert@cs.glasgow.ac.UK (Gilbert Cockton) (08/25/88)
From: Gilbert Cockton <gilbert%cs.glasgow.ac.uk@NSS.Cs.Ucl.AC.UK> Date: Tue, 23 Aug 88 05:54 EDT To: ailist@ai.ai.mit.edu Subject: Re: Dual encoding, propostional memory and... In reply to Pat Hayes last posting >Yes, but much of this debate has been between psychologists, and so has little >relevance to the issues we are discussing here. [psychologist's definition of different defined] >That's not what the AI modeller means by `different', though. >it isn't at all obvious that different behavior means different >representations (though it certainly suggests different implementations). How can we talk about representation and implementation being different in the human mind. Are the two different in Physics, Physiology, Neurobiology .... And why should AI and psychology differ here? Aren't they adressing the same nature? I'm sorry, but I for one can see how these categories from software design apply to human information processing. Somewhere or other, some neurotransmitters change, but I can't see how we can talk convincingly about this physiological implementation having any corresponding representation except itself. Representation and implementation concern the design of artefacts, not the structure of nature. AI systems, as artefacts, must make these distinctions. But in the debate over forms of human memory, we are debating nature, not artefact. Category mistake. >It seems reasonable to conclude that these facts that they >know are somehow encoded in their heads, ie a change of knowledge-state is a >change of physical state. Thats all the trickery involved in talking about >`representation', or being concerned with how knowledge is encoded. I would call this implementation again (my use of the word 'encoding' was deliberately 'tongue in cheek' :-u). I do not accept the need for talk of representation. Surely what we are interested in are good models for physical neurophysiological processes? Computation may be such a model, but it must await the data. Again, I am talking about encoding. Mental representations or models are a cognitive engineering tool which give us a handle on learning and understanding problems. They ae a conative convenience, relevant to action in the world. They are not a scientific tool, relevant to a convincing modelling of the mental world. >what alternative account would you suggest for describing, for example, >whatever it is that we are doing sending these messages to one another? I wouldn't attempt anything beyond the literary accounts of psychologists. There is a reasonable body of experimental evidence, but none of it allows us to postulate anything definite about computational structures. I can't see how anyone could throw up a computational structure, given our present knowledge, and hope to be convincing. Anderson's work is interesting, but he is forced to ignore arguments for episodic or iconic memory because they suggest nothing sensible in computational terms which would be consistent with the evidence for long term memory of a non-semantic, non-propositional form. Computer modelling is far more totalitarian than literary accounts. Unreasonable restrictions on intellectual freedom result. Worse still, far too many cognitive scientists confuse the inner loop detail of computation with increased accuracy. Detailed inaccuracy is actually worse than vague inaccuracy. Sure computation forces you to answer questions which would otherwise be left to the future. However, having the barrel of a LISP interpreter pointing at your head is no greater guarantee of accuracy than having the barrel of a revolver pointing at your head. Whilst computationalists boast about their bravado in facing the compiler, I for one think it a waste of time to be forced to answer unanswerable questions by an inanimate LISP interpreter. At least human colleagues have the decency to change the subject :-) >If people who attack AI or the Computational Paradigm, simultaneously tell me >that PDP networks are the answer I don't. I don't believe either the symbolic or the PDP approach. I have seen successes for both, but am not well enough read on PDP to know it's failings. All the talk of PDP was a little tease, recalling the symbolic camp's criticism that a PDP network is not a representation. We certainly cannot imagine what is going on in a massively parallel network, well not with any accuracy. Despite our inability to say EXACTLY what is going on inside, we can see that systems such as WISARD have 'worked' according to its design criteria. PDP does not accurately model human action, but it gets some low level learning done quite well, even on task requiring what AI people call intelligence (e.g. spotting the apple under teddy's bottom). >Go back and (re)read that old 1969 paper CAREFULLY, Ah, so that's the secret of hermeneutics ;-]