In complexity and AI I explained an advantage AI learning has over scientific investigation. Science is looking for explanations that rise to the level of the general -- a theory. AI learns behaviorally -- empirically -- without having to explain why any particular datum behaves oddly. It simply has to learn it. AI is a technology; as a technology it has a task to achieve. Explaining how it learned or became what it is is beyond its pay grade. But science has to incorporate the odd behavior into its theoretical explanation. Since the real world is unfathomably complex in its interactions, influenced by multiple internal and external pressures, odd noises abound. That's a problem for scientific theory. Tech doesn't have or need a theory in order to learn or achieve its task.
The English language, for example, has lots of Norman French words as well as Viking Norse words to complicate the vocabulary of its Anglo-Saxon base. These were not natural evolutions within the language. They're in English because of historical invasions that are entirely unrelated to language. But English also has mistakes that speakers make that spread and survive, for example back-formations ("pea" is derived from the old word "pease" not from a plural "peas", "edit" a verb invention from "editor", "burgle" from "burglar"). If you want a theoretical explanation for the language faculty, you've got to know which data come from which pressure -- invasions or human mistakes or language internal shifts. But a behavioral-empirical learner doesn't care where these pressures originated. It's got no theory to prove or promote. All it cares about is producing the behavior, regardless of its source or history.
To state the irony clearly: a general theory of a phenomenon should facilitate learning and understanding of that phenomenon. Like a map, it should show the way in any direction. Yet for an empirical-behavioral learning program, a theory would be an encumbrance since such a theory wouldn't predict the pressures external to the phenomenon. The map that shows all and only the roads, doesn't warn you about the hills or the quality of the road or the traffic congestion. Behavioral learning deals with these directly, on the ground. That's why guerilla warfare can be so effective against large-scale military intelligence.
When engineers gloat over having succeeded in learning English while linguistic science still can't explain how children learn the language, we should remember that the goals of science and the goals of technology are completely different. Technology doesn't need an explanation, much less a theory. Do AI engineers know how AIs learn particulars? They know the tools the AI uses -- the engineers built them! To learn more than that, they have to experiment and test their own creation. And the AIs themselves don't know any better either.
Notice that this is also true of humans. Children don't know how they learn their native language either. Kids learn their language without even knowing that they're learning it or thinking about it. Learning a language by studying its explanation -- that's what a grammar book is -- turns out to be a difficult task. And if the grammar book elevated itself from an explanation of the language to a general theory of language, it would be pretty much useless to the language learner.
The goal of technology is to achieve a task for a consumer -- an individual consumer, a corporate consumer or a military or other government consumer. It doesn't matter how it achieves the task as long as it works. That's the shallowness of technology. It needs no theory, just practical learning means. Science takes on a harder task.
The engineer who designed the bicycle assumed that turning the handlebars would successfully turn the bicycle. But any bicyclist knows that if you don't also lean in the direction of the turn, you'll fall over. In other words, the machine worked, but the engineer who made it didn't quite know exactly why or how. He provided a necessary element of turning, but not a sufficient. In that sense, he got lucky. Programmers know this only too well. And complex program has to be not only reviewed, but given a test run, before being implemented.
If you ask an AI about its learning, it'll tell you what the literature says about its learning, which may be far from how the AI actually learned. It has no more introspection than a child has -- or an adult for that matter -- about how it learnt to speak or walk or any of its basic learning accomplishments. Explaining those is the task of science.
The goal of science is not to accomplish a task in the world or change the world in any way. It's to understand the world as it is and explain it. And that includes the technologies that engineers have introduced into the world. They too need scientists to experiment, test and explain and understand.
Here's a presentation I gave to the International Linguistics Association titled "Why neural networks succeeded where generative linguistics failed: The syntactic and semantic limits of neural network behaviorism (Yoneda lemma limit)" based on the post complexity and AI. It addresses the extraordinary ironies in the development of linguistics and AI. In 1956, Chomsky showed that simple empirical-behavioral learning was too shallow to account for the productivity of language. It turns out that that shallowness is exactly what facilitates AIs' empirical-behavioral learning, a shallowness that is not available to the linguist scientist.
It also explains the difference between tech and science; innate learning (like how to walk or talk) and non innate productive learning (like learning how to ride a bike or learn math),
and why for science it's just as important to describe what humans cannot do as a consequence of our evolution as a species, a matter of no concern at all for tech since its goal is to achieve tasks, not limit itself.
No comments:
Post a Comment