Which is newer?
Meanings are hard which means AI is hard.
In October, I purchased Robert Skidelsky’s Money and Government. In November, I purchased David Graeber’s Debt. Which book is newer? It seems obvious that the more recently purchased book is newer. However, Debt was published in 2014 and Money and Government was published only in September. Does that change which is the newer book?
What if, today, I procure a first edition copy of Noam Chomsky’s Aspects of the Theory of Syntax (published in 1965). Now, which is my newest book?
Publishers sometimes issue reprints. A reprint, according to Wikipedia, is “a re-publication of material that has already been previously published,” with no modifications to the content—if the content is modified it’s called a new edition.
Suppose MIT Press, the publishers of the above Chomsky book decide to issue a reprint and I buy the reprint too. Is the reprint my newest book? Would this change if I’d first purchased the reprint and then the first edition?
ebooks can further complicate the situation. Suppose I download two PDFs of Dean Baker’s Rigged (published 2016). Which amongst these two is the newer copy?
Leaving aside the virtual, Dean Baker’s book is published under a Creative Common’s license (not that it matters but the same license as this post). This means it’s legal for me to print a copy. Suppose I have a copy of Rigged printed and nicely bound. Is that my newest book? Even though it was published in 2016? What if I buy a paperback version of the book? What happens then?
Now let’s look at the views of two cognitive philosophers on the same topic. Where I’ve briefly discussed an adjective, newer, we’ll see Chomsky on nouns and Fodor on verbs. Fodor is especially funny.
Noam Chomsky, Institute Professor & Professor of Linguistics emeritus, philosopher, political-activist, and all-around nice guy, points out that we already know these things and we don’t have to be taught them. In fact, we couldn’t be taught them because any attempt to teach someone these “facts” would only confuse them. Imagine including these above examples as part of a cross-cultural curriculum. First, the teachers would look at you like you were crazy. And even if you managed to convince the teachers, the students would walk out in just a few minutes.
In his 2000 book, New Horizons in the Study of Language and Mind, Chapter 2, “Explaining language use”, Chomsky gives some great examples of things we all know but perhaps don’t realize we know it. “Whether something is properly described as a desk, rather than a table or a hard bed, depends on its designer’s intentions… among other things.” We can refer to objects in a concrete manner “the book weighs five pounds” or abstractly “who wrote the book?” or both simultaneously “the book he is writing will weigh five pounds if it is ever published”. Another example: “Peter and Mary [who] are equidistant from the surface — [of a house, with] Peter inside and Mary outside — Peter is not near the house, but Mary might be, depending on current conditions for nearness.”
Chomsky’s third example leads us to the difficulty we have with even simple definitions: “if John is painting the house brown, then he is applying paint to its exterior surface, not its interior”. Chomsky also points out that the “fact that a brown house has a brown exterior, not interior, appears to be a language universal, holding of ‘container’ words of a broad category… [like] box, airplane, igloo, etc”.
Painting something seems fairly obvious. How would you define the meaning of the transitive verb “paint”? The philosopher Jerry Fodor attempted to do this and failed. His reasons are worth examining at some length primarily because they’re so funny. The quotations below are taken from his 1983 book Representations, Chapter 10, “The Present Status of the Innateness Controversy”.
Let’s review a few of Fodor’s definitions of “paint” that don’t work:
- “x covers the surface of y with paint”. This doesn’t work because “you could have painted the wall even if there were a little tiny bit up there in the top left hand corner that you’d missed, hence failed to cover with paint… it might be that you’d painted the wall… not very well; but, nevertheless, you’d have painted it.”
This is a good counterargument but Fodor doesn’t stop here; no, he gets funnier still. “What’s wrong with the definition” of equating ‘to cover’ and ‘to paint’ is that “there are clear counterexamples” of things that describe “to cover” but are not “to paint”. For example, Fodor says, “to start with a fairly crude point, consider the case where the paint factory explodes and covers the spectators with paint. This may be good fun, but it is not a case of the paint factory… painting the spectators.”
- Clearly, the accidental coverage of a surface with paint does not constitute painting it. Fodor then amends the definition to say, “x paints y iff x is an agent and x covers the surface of y with paint”. (iff is a contraction of if and only if.) An agent is someone (perhaps including an artificially intelligent entity) that does something purposefully.
“The trouble is,” says Fodor, “that this definition doesn’t work either.
“Consider that Michelangelo, though an agent, was not a house painter. In particular, when he covered the ceiling of the Sistine Chapel with paint, he was not painting the ceiling; he was painting a picture on the ceiling.”
- Okay, what’s next? Fodor continues, “it’s not good enough that you cover the y with paint… [or] that you cover the y with paint intentionally (though that’s necessary since, if you knock over the paint bucket, …, you have not thereby painted the floor). It’s got to be that when you cover y with paint, what you primarily have in mind to do… is that y should be covered with paint in consequence of your activity.
“Anyhow, this definition doesn’t work either. For consider that when Michelangelo dipped his brush into Cerulian Blue, he thereby covered the surface of his brush with paint and did so with the primary intention that his brush should be covered with paint in consequence of his having so dipped it. But MICHELANGELO WAS NOT, …, PAINTING HIS PAINTBRUSH. (He was just putting paint on his paintbrush.)” [emphases in original.]
After these three arguments Fodor gives up and admits that he doesn’t “know where we go from here”.
Whither Artificial Intelligence?
What does this mean for AI? The goals of AI are usually to have a computer simulate human behavior as closely as possible. (The goals of cognitive science are to understand human cognition like a nuclear physicist wishes to understand elementary particles or a primatologist wishes to understand monkeys. Scientists don’t know if any additional practical implications will follow their research, and assume knowledge for its own sake is a good thing.)
Computer scientist, Robert Munro, recently had an elaborate, detailed article in his blog, provocatively titled “Diversity in AI is not your problem, it’s hers”. He discusses how the modern Machine Learning systems which primarily use statistical models (and zero understanding) have failed to identify “hers” as a feminine pronoun but have no problem understanding “his” as one. It’s an interesting aspect of the English language, one I’d never before noticed. However, what Munro doesn’t explain or get into are the kinds of problems Fodor and Chomsky have been discussing for quite a while now. And these problems are much deeper than the ones Munro describes (which are also very interesting).
Patrick J. Hayes
There are computer scientists that recognize this. In 1987, a book titled, The Robot’s Dilemma, was published, containing essays by both cognitive philosophers and AI researchers with a specific focus on The Frame Problem. The Frame Problem was first formulated in 1969 by two pioneering researchers named John McCarthy (who coined the term Artificial Intelligence) and Patrick Hayes (pictured above). Their paper was titled, Some Philosophical Problems from the Standpoint of Artificial Intelligence. (The Wikipedia page is pretty good too.)
In The Robot’s Dilemma, Patrick J. Hayes has an essay titled, “What the Frame Problem Is and Isn’t” which describes the recognition and early synergy between philosophy, computer science, and cognitive science. Below is an excerpt from Hayes’s essay and it’s also a good place to end this post (emphases added):
Elsewhere Fodor refers to Hume’s problem, of identifying “natural”… relations. These are, of course, the primitive symbols of the system’s axiomatized beliefs. Fodor is quite correct here: AI is indeed trying to locate these Humean relations….. In his characteristic way, Fodor simply asserts that “we” have no idea what these relations might be. Well, again, the answer from AI is that we are working on it, and we have got somewhere, but there’s a long way to go.
December 7, 2019, Happy 91st birthday, Noam Chomsky!