Previous entry : Martin's follow-up

Next entry : Yesterday's RSS Meetup

July 12, 2004

The unbearable inevitability of discretization (2/2)

Here I am with part 2 of my attempt to undemonize the eventuality of an intelligent Semantic Web. BTW, you really should read part 1 first. Take note that I really consider the intelligent Semantic web machine as the system comprised by all the data, transport mechanisms, and agents (interactive or not) of the web.

We are not that hot (read complex, subtle and nuanced)

This is the part where I stand on the unpopular side of mathematical determinism. Yes of course, our intellect is powered by mechanics that benefit from a couple of millions of years of natural selection, whereas our silicon counterpart is only fifty years old, at best. And yes, this gives our nervous system some obscure capabilities to leverage on. But, there are significant parts of our psyche that we can actually modelize pretty satisfyingly. By satisfyingly, I mean at least as satisfyingly as the accuracy of the current meteorological models. Isn't it Pavlov dearest?

You don't need to be a neuropsychologist to understand that we (humans) follow a lot of very simple cognitive patterns (read near first-order logic algorithms). And that these cognitive patterns account for a strong proportion of our daily activities (sex sells, just face it). This (somewhat) neat functional compartmentalization of our intellectual mechanics becomes very obvious in case of damage to the nervous tissues : bugs arise and often unveil the specialization of the affected areas. And some of the most complex specializations are often surprisingly localized, suggesting the potential modelizability of their underlying machinery. And since these specializations include hard-coded semantic processor as well, one can rightly imagine that there exist algorithms that can efficiently reproduce the behavior of the corresponding organic systemics. Therefore, it is very hard to believe we will never, ever have synthetic brains rivaling the original. If you haven't already done so, please read Oliver Sacks's enlightening The Man Who Mistook His Wife for a Hat for an entertaining journey into our not so complex and nuanced self.

We also largely overestimate our capacity to mentally resolve subtle language problems. For example, as far as semiosis goes, we do not handle communication errors in an infinitely complex way. far from that. Social semiotic crashes occur between humans very frequently. I do not need to stretch very far for examples : this very debate and the way it was conducted at the last RSS meetup is a convincing illustration. People meeting for the first time (some of them anyways) were trying to discuss a subject which basic notions are very vague. After a clear online, textual rendition of the conversation, we are still asking ourselves if we talk about the same things (see Martin's comment on the translation I made of his own post). And just think of how many times per day you will hear or say that's not what I meant. It is then a little pretentious to say that AI is irrelevant because it cannot recover perfectly from communication error. Fail-fast is, in both the human and semantic web worlds, a best practice anyways.

Another personal example of our inability for reliable and lossless semiosis is the game I played with (some of) my French litt teachers. In primary school, literature classes essentially consisted of first-degree text analysis, like What is the color of Napoleon's white horse?. In high school, however, we were introduced to the marvelous world of poetry and figures of speech. I first realized that the game was not about knowing what a given text meant anymore, but rather to try to guess what the teacher believed the text meant. Then, it became clear to me that as long as you could structure your ideas following very simple rhetoric rules, you could convince your teacher and the rest of the class on whatever meaning you wanted for a given text! I remember of one oral presentation where I successfully defended that 1984 is exclusively a romantic love story. To me, that was the end of the idea that figures of speech could be used to produce anything but not so meaningful entertainment. I was later introduced to Jacques Derrida's theory on semantic deconstruction, which definitely ranked the argument for language formalization being the opposite of language expressiveness amongst the ugliest sophisms. In the same vein, it has been argued that machines will never cope with an ever-changing semiotic consensus. Correct me if I am wrong but there is no "objective" semiotic consensus outside collective indices like Wikipedia and PageRank. And what is objectivity anyways. Didn't it die three thousand years years ago?

I always preferred business computing imperatives over liberal arts computing misconceptions. When I read my idea of semantic web is if I can look for a poem that uses a metaphor of bird as freedom, and get back poems that have bird as metaphor for freedom it really feels weird inside of me. Humans themselves are unable to consistently interpret poems and other malformed free text the way they were conceived by the authors. I mean, just what, exactly, is I Am the Walrus about? On the other hand, business folks know what they want and are able to formulate a sound query. Not surprisingly, the business world has seen the first successful implementations of the semantic web's building blocks. We are generally unable to make sense, and that's computing's biggest obstacle.

I won't cover more extensively that part of my argument because all in all, you first have to meditate your perception of your own mental efficiency. And because it always involves emergency ego landings, it takes years, literally. If you are still stuck with the Renaissance's utopias of love, beauty and truth, and if you give credit to the so-called empiricism of Soul, Will, Freedom and Consciousness, no amount of logic and rhetoric can convince you. You are in an intellectual no man's land. Seriously, it's just that even from a supposedly scientifico-materialistic society, 99.99% of the anti-AI arguments (yeah, maybe a little less than that) simply are mystico-delirious rewordings of the major semitic religious gnosis.

AI as a practical technological orientation rather than a punctual objective

Intelligence is like love. Everybody uses the term, nobody knows what it means. And just like nobody wants to know love is just an opaque wrapping for sex, nobody wants to know human intelligence is a purely material organic buildup. Nobody wants to know one day we will loose the intelligence championship to something we have created. We maintain a totally artificial sense of grandeur of the human nature and it is really slowing us down in many ways.

The Turing test is the epitome of the victim of humanism. As soon as we feared that machines could potentially pass it, we simply discarded both the test and the then generally accepted definition for intelligence. With pseudo defeats such as the chinese room, we successfully obfuscated the very notion of intelligence. It is obvious that a formal definition implies a model and that a model implies a mechanization. And since we can't suffer the eventuality of AI, we say that the Turing test merely indicates successful intelligence mimetism. Ironically, we are still very proud of how much we score in IQ tests that are exactly of the same type as Turing's. Yet we don't say writing an IQ test is about mimicking intelligence! Silly romantic weasels, we are! We will not be satisfied with seemingly intelligent outputs for given inputs, which is the definition of an intelligent machine : we will only accept human intelligence, period. The chinese room is our collective commitment not to recognize intelligence unless it comes from a bloody organic human brain, which is in essence a circular argument : something is human because it is intelligent, and something is intelligent because it is human.

Let me give you my definition for intelligence : intelligence is a relative metric for cognitive efficiency and universality. That means nothing is simply intelligent by itself : something is more or less intelligent than something else for a set of given tasks. It means that for natural language parsing, computers are currently less intelligent than humans, but more intelligent than a swiss army knife. It also means computers are more intelligent at natural language parsing now than they were 10 years ago. There is no absolute intelligence universality, either. Mechanical cognitive universality is a evolutionary process, not a punctual objective that we will one day attain.

I do personally believe computers will undoubtedly show greater intelligence than humans the day they will outperform them at the game of Life, and I believe it will some day happen. But in the meantime, passing the Turing test still remains an important milestone and for software to pass the Turing test means it's more intelligent than its opponent, as far as IQ goes.

Evolutive efficiency also applies to the Semantic Web. Luckily for us, it benefits from two distinct evolutionary avenues. It indeed gains effectiveness both from cleverer agents and from semiotically-complete ontology representation formats (relational databases, XML/RDF, OWL, UML, etc.). Therefore, with some site correctly implementing the Semantic Web-enabling technologies, one is right to argue that the Web already shows some signs of semantic intelligence.

Discretization is the fundamental mechanism behind any form of cognition. Solve et coagula-based computing rules!

Discretization is about appropriately lowering the level of detail. It is indeed all about contextually discarding, discriminating, weighing information. You don't need to take relativity into account when you are assembling your IKEA chair. When you are performing a mental forecast for the interest rate for two years, you most probably do not want to handle the eventuality of a butterfly landing on the Eiffel Tower within that timeframe. That's because unless quantum computers tackle your problem differently (and you have one handy), trying to compute, with your brain or PC, any form of continuum is the cognitive equivalent of Zeno's paradoxes. You will get nothing out of it, trust me on this.

Discretization's superstar is calculus. To better understand and be able to perform operations on otherwise complex mathematical objects (resp. the curve and the arbitrary surface), we gave ourselves simple linear models (resp. the line and the plane). It so happens that since calculus, we discovered we understand the world around us by breaking it down. Simple models can surprisingly yield extremely complex knowledge acquisition. From abstract, simple mathematical constructs, we came up with nuclear fusion, microwave ovens, and magnetic resonance imaging. Clay Shirky's main argument against the semantic web is that syllogisms don't work well in the real world, because most of the data we use is not amenable to such effortless recombination. Discretization is the anti-thesis to just that.

The bet of maths in general, and AI and the semantic web in particular, is that complex knowledge arises from simple models combined with simple algorithms, and discretization's success is the proof of concept. I'm not alone on that line of thought, just have a look at some of the work of my favorite junk (yes, junk, even pre-empirical) scientists on the subject : Gödel, Escher, Bach from Douglas Hofstadter and A New Kind of Science from Stephen Wolfram.

Semiotic, our very recent will to discretely modelize modelization, is in my opinion, the noblest intellectual endeavor of our times. The concurrent models of semiotics' founders Saussure and Peirce are the ancestors of today's RDF, and if we look at what we got from older mathematical models, in physics and chemistry for example, one can only hope the most exciting things are still to come in the semiotic field, and by extension, from the semantic web.

Discretization is our rational march towards relevance.

Getting over the AI skeptic romantic FUD, with the semantic web as the most exciting thing since Jesus' resurrection

If you say that a semantic web malformed-query interpreter hyperGoogle that would pass the Turing test is trivial and would only impress a novice. I'd say I'd be happy to be a novice amongst folks like Jon Udell, Stefano Mazzocchi, Tim Bray, Rendall Clark, Tim Berners-Lee, etc., etc., etc., etc., etc., etc., etc., etc., etc. I would also ask you what would you be impressed with? While you think about it, can we at least agree on the usefulness of some of the W3C's recommendations?

Tomorrow is RSS meetup day. If I feel I can safely say two consecutive words without risking my life, I would like to share with the attendants the simple, open source software ideas I, along with some friends, have been nurturing for a while now. I will be showing these things off here as soon as I get them down textually.

Posted by Vincent-Olivier Arsenault at July 12, 2004 08:10 PM