Reflection on extreme representationalist perspective

Latest revision as of 04:56, 12 February 2009

I am focusing on the Schank/Abelson ('Abelson' for short) and the Newell/Simon ('Simon' for short), though before digging in I'd like to give a word to the textbook, "Human Reasoning...", which seems to be existing in a plane independent of the rest of the course, at least because it has gone entirely unmentioned. I am enjoying it, and greatly appreciating it, because I can see that it is challenging me. Berkeley's Cog Sci program, despite Stuart Russell's best efforts, left me in the 'eliminativist' camp--"reasoning rarely happens in real life, and mainly in institutional contexts...fast and frugal algorithms...constraints of time and energy...evolutionary psychologists..." I am looking forward to forcing myself to challenge that view in the manner proposed by the authors. I haven't been following everything 100%, but I also haven't been giving it 100%. I'm sure that I'll come through the book being able to articulate either why I end up agreeing or disagreeing, but I don't know what role you intend it to take in the course, so let me know if I should be engaging more closely with it.

As for Simon and Abelson, what struck me most is that they are (or were) considered to be in the same camp of AI. Though the confines Abelson had in mind for the structure of thought bear much resemblance to Simon's, the latter gives less impression of having seriously considered the uncompromising complexity of the domain. This winter I read Skinner's Beyond Freedom and Dignity, entirely prepared to find that the behaviorism I had been told about was just a straw man built to pose weakly against its obviously superior successors. Specifically, I was expecting Skinner to instead make the very reasonable claim that there may be mental states, but we should proceed as if there aren't. What a surprise to find that his account of cognition was just as unreasonable as I had been taught. I feel the same way about Simon and Newell's approach to intelligence. I would be happy to concede, and would have expected to hear, that search offers a useful toy model of intelligence. Instead I read that search is being offered not just as a model of intelligence but as sufficient for creating it.

In my undergrad, the preoccupation with search killed any potential personal interest in AI. At this point I can't read papers in this tradition without my mind wandering through the possibilities offered by one key counterfactual: "What if Turing had played Go instead of Chess?" I like to think that the field would have never gotten lost in its search for intelligence in search, or so rooted in trees. We would instead have unleashed all of these important thinkers directly on the hardest and most central of Hume's three directions: Similarity.

I am grateful to all of these researchers for how easy they are to pin down, in this respect they are scientists. Abelson betrays his ties to GOFAI with his equally responsible procedure, leaving me with a clearly specified assumption to take special exception to. I couldn't say the same of Merleau-Ponty, whose attempts could never have led so immediately to an actual implementation. By contrast, my introduction to the Restaurant Script was through algorithms and structures for representing it, sterilized of all the wonderful meaning, thoughtfulness and complexity of consideration offered in these chapters (which I am only just seeing, leaving me a bit sad that it took so long to vindicate Abelson).

The clearly specified assumption I was able to take exception to was the second axiom of Conceptual Dependence Theory: "Any information in a sentence that is implicit must be made explicit in the representation of the meaning of that sentence" Though I am tempted to feel much satisfaction in my ability to immediately identify its weakness, I can instead only find gratitude to Abelson and Schank for providing clear falsifiable axioms on which to build an impressive account. It is much easier to destroy a tower than build one, and I can't say that I could replace his theory, as old as it is, with anything nearly so coherent and accounting for nearly as much of the data.

I'd like to say that Abelson's much more serious engagement with the real dirty subtleties of thought allowed the flexibility that has made his lab with Sussman at MIT a center for countless exciting and unconventional projects, like the various global-to-local compilers and other impressive, incredibly decentralized approaches to emergence and intelligence. By contrast, insofar as Minsky's recent book can be taken to represent the contemporary perspective on Newell and Simon's school, the classic computational account of intelligence is going to look the same dead as it did alive.