Conversations with Joshua
Summary: The term __ldq__Artificial Intelligence__rdq__ is so abused it's impossible to properly define what it is anymore. Disenchanted takes on the problem by splitting it up into five categories, with the first__em__conversational machines__em__being the topic of this first article in the series.
Marvin Minsky says we shouldn't intimidate ourselves by admiration of our Beethovens and Einsteins, and that there isn't much difference between ordinary thought and highly creative thought. If we could just somehow simulate the basic process on a computer, then in theory that computer should be capable of creativity as well. “Artificial” intelligence is our objective—a machine that thinks just like humans do. This field is broad enough that I've split it up into five categories that I'll be discussing in five articles; with one each month. What I'll discuss will not only include how intelligence works and could be simulated, but what purposes mankind will put it to and the impacts it'll have on humanity's own way of thinking. (Plus, you'll also see why I've chosen to put quote-marks around the word “Artificial.”) This article is the first in the series and will talk about the first and best known category of AI—the conversational, or “chatty”, machines.
Wouldn't you rather play a good game of chess?
Conversational AI begins with the prejudice that language is not just the best way to judge sentience, but is also the best way for the computer to percieve the world. In fact, the language prejudice is embodied in the de-facto test of an AI's worth: the Turing Test.
It works like this: Put a human in one room, a computer in another, and a human Interrogator in a third room. They're all connected together by computers and copies of AOL Instant Messenger (we've modernized the tools). The Interrogator isn't told which of the two contestants is the computer, merely that one is called ‘A’ and the other is called ‘B’. The Interrogator procedes to ask questions of the two contestents until he figures out which one must be the computer. The computer, of course, is pretending to be a human.
Let's say that the Interrogator asks “What's the weather like today?” and gets the following responses from the two contestants.
Player A says: “It's hotter today than it was yesterday, and my clothes are sticking to my skin.”
Player B says: “I hate it. I'm sweating like a pig and its making me too uncomfortable to sleep at night.”
The Interrogator might already be suspecting that Player A is the robot, since it merely relayed factual information (and unimaginatively, at that), while Player B actually expressed feelings. The comment about “my clothes are sticking to my skin” could have been pulled out of a table of responses keyed to the topic of the question—responses thought of in advance by the programmer.
The Turing Test is rather inexact, though, since a real but unimaginative human could potentially lose. For this reason most contests subject the machines to more than one judge. In the case of the Loebner prize, for example, the AI who fools the most judges “wins” (the Loebner prize actually has two levels of “winning”—$2,000 for the one that fools the most judges, and a real prize of $100,000 if the computer fools more than half of the judges).
It's of interest to note that the original form of the Turing test, as described by Alan Turing himself, doesn't involve any machine candidates at all, but has a man pretending to be a woman instead (suddenly, AOL Instant Messenger is relevant, again).
The second forgotten detail of Turing's original test is that the Interrogator isn't supposed to know he's looking for a computer at all.
We call it Voight-Kompf, for short.
If the Interrogator knows he's trying to find the robot then it makes it a lot harder for the computer to fool him, perhaps unfairly harder. After all, we don't interrogate each other in routine life. If a computer is simply looking to blend in with a crowd, then there are programs that have already succeeded.
While the original Turing Test gives the computer such an opportunity, its perversions of late have lead to non-constructive trick questions as part of the Interrogator's repetoir. “What does the letter ‘M’ look like upside-down?” may be impossible for a computer to answer without a programmer anticipating it, but its failure to answer that question doesn't say anything about its worthyness as a mind. Most humans blind from birth can't answer that question, either.
A better test of a robot's intelligence may be to lock it in a room with a basket of household cleaning products, a book on explosives, and the instruction to “get out.”
This kind of test would not only check for problem solving intelligence, but also its self awareness. For a robot that solves the problem of making the explosive, but uses its own body and battery to hold and detonate the charge will fail as a human-like thinker. The machine certainly solved the problem from the mathematical sense, but destroyed itself in the process. Why? Because animals and machines that aren't self-aware will inventory their own bodies as a disposable resource, not making the cognitive link between the body and the source of its own thoughts. Although it might alter the outcome of the test, you can't just give it a new directive (“do not destroy this unit”) and expect it to be the same.
Hmm... the unfreezing process seems to have left me with no internal monologue!
We don't need real rooms, explosives and robots to do these kinds of tests, for all could be simulated on a computer. The language prejudice, however, might come not from the expense or difficulty of making physical simulations, but from the internal monologue—the “voice inside our head.” We “think” in our native tongue, and so believe a computer should, too.
But it's not altogether clear that language skills are inseparable from intelligence, and again we can look to the capable intelligences of our blind, deaf, and brain-damaged for indicators.
Aphasia is a condition brought on by physical damage to the “language organ” of the brain, such as from a stroke or a wound. Since language skills seem to be spread across the brain (with hearing, speaking, reading and writing taking place in different parts) the severity of aphasia will depend on where the brain injury occured. Broca's aphasia—for example—affects the ability to generate words (the patient may say “Walk dog” to mean anything from “I want to walk the dog” to “you take the dog for a walk”), while those afflicted with Wernicke's aphasia will speak in long and poorly constructed sentences with lots of nonsense words (“You know that smoodle pinkered and that I want to get him round and take care of him like you want before” means “The dog wants to go out, so I'll take him for a walk”). Global aphasia victims may lose all understanding of language completely.
Yet despite losing or confusing the gift of the tongue, aphasia victims do not become stupid. In fact they're painfully aware of their problem and are horribly frustrated with it. And they can still solve problems, deal with abstract concepts, recognize patterns, make cognitive leaps and behave as self-aware beings. Could this suggest that language, while a characteristic of intelligence, is not critical for it?
Must... think... in Russian. Think... in Russian.
You can argue that aphasia doesn't discount the possibility that the victim may still have a linguistic internal monologue and is simply unable to express it. A leading theory of thought is that we don't actually think in words anyway, but in structures of associations. It's our own sense of “listening to ourselves think” that's fooling us because the associations to our memories of spoken words are so strong and “well trodden.” Imagine the President of the United States in a clown suit. Have you actully seen him in a clown suit before? If not, how did you manage to imagine that scene? Is there a copy of Photoshop inside your brain that is frantically manipulating bitmaps as fast as you can think?
What could more easily explain that ability is if your brain has not actually “painted a picture” of George Bush in a clown suit (no offense intended, Mr. Bush, we just picked the example because we're pretty sure nobody's seen you with a big rubber nose) but merely manipulated sequences of associations. We remember seeing a clown before, our brains' visual center was able to identify all the sub-objects and attributes of the face (eyes, nose, mouth, ears, color, etc.), and all we had to do was think of the attributes that represent our Fine Leader's face combined with a few extras that we remember from elsewhere.
We think there's a crisp photograph inside our heads, but is there? Could you wire your brain up to a computer that could measure synaptic activity and build a picture on the screen by finding the grid of neurons that presumably represent all the pixels of the picture? No, you could not.
You're being a bit brief, perhaps you could go into detail.
To get an idea of what we mean by “structures of associations”, we can examine one of the many ways you can write a conversational AI. Returning to the language prejudice, we want a routine that can parse an English sentence and figure out its meaning. “George is a cat”, for example, can be picked apart into a very simple knowledge tree.
- George
- is a cat
The computer has focused on “is a,” which it knows is a type of relationship. Elsewhere in the machine's memory is another knowledge tree that describes what the “is a” means in template form.
- Entity identifier
- 5: is a entity type
Matching the parsed sentence with the template tells it that “George” is an Entity identifier (a name), and that “cat” is an entity type. The number 5 represents a confidence level—it's halfway sure that any sentence matching this template is identifying an entity (George) and then giving it a type (cat). Going further on, the programmer might have also told it a few things about the word “cat.”
- Entity type: cat
- 8: is a animal
- 10: is a self-directing entity
- 10: has claws
- 10: has fur
- 4: is a sports car
- 10: is a deterministic entity
- 10: also known as Jaguar
- 10: has engine
- 8: is a animal
Notice here that the computer is pretty sure (confidence is 8) that “cat” on its own nearly always refers to an animal, but is also “aware” (just less confident) that we might be using slang for a certain British automobile instead. If we'd said more in our opening sentence, like “George is a cat, his claws are sharp”, then it could have matched the attribute of claws—boosting its confidence up to 10. Then it'd be really sure we're talking about an animal, specifically one that has fur and makes its own decisions. (Do you know a cat that doesn't have a mind of its own?)
It doesn't matter what the words are that you use to label each node of the tree, we could use numbers instead to represent concepts such as a “cat” and then associate them with memories of words (analogous to our memories of word sounds and word shapes). But these knowledge trees do not represent understanding, they're merely a convenient data-structure upon which we can hang rules, so our “chat-bot” can at least produce replies that make sense.
The job of formulating a response is a little trickier, but again we can use a tree, this time as a decision-making structure.
- If subject is a self-directing entity, then
- 10: say “How is (entity pronoun) doing, these days?”
- 5: say “What's (entity identifier) like?”
- 1: say “I see, tell me more.”
- If subject is a deterministic entity, then
- 10: say “How much did it cost you?”
- 5: say “Where did (entity identifier) come from?”
- 1: say “I'm not sure I understand what you're telling me.”
What this is saying is that if the computer is really sure (confidence is 10) that the subject is a self-directing entity, then it'll ask a question that assumes so (“How is he doing, these days?”—asuming another knowledge tree matches “George” with a male pronoun), but if it's less sure (confidence is between 5 and 10), then it'll play it safe with a more generic response.
The more sophisticated the knowledge and decision making trees are, the more “intelligent” the computer will seem. A clever programmer will have it ask questions that would resolve a lack of confidence, to make it seem as if the computer is able to intuit what we're talking about. And these examples of knowledge trees, although very simple, are meant to give you an idea of how a structure (tree) of associations (to memories of sensations) might work in your own mind. It's certainly easier to solve our earlier thought problem, now.
- George The Clown
- has a concerned frown
- has a big red nose
- has a medium build
- has a pair of funny shoes
But least we imagine that given a sophisticated enough set of knowledge and decision-making trees, then the computer could begin to solve problems (or even have emotions), remember that this kind of program is merely reactionary. No thought processes occur until the human has finished typing and presses the ENTER key, and then all it does is follow a set of rules that a programmer thought of beforehand. It's not solving problems, it's merely matching the input with solutions that have already been thought of by the programmer. If trees of associations are how we store knowledge in our own minds, then all the computer is doing is borrowing a couple of tricks—but not The Trick.
These types of programs, where knowledge and decisions are represented in trees, are better known as expert systems, and they're great for diagnosing car trouble.
Open the Pod-bay doors, please, HAL
Expert systems and a rudimentary ability break apart sentences to find meanings give us a practical application for the conversational AI: user interfaces. We've come a long way since Leisure Suit Larry, but combined with mature voice recognition (which itself makes use of another AI technology—the neural net—which we'll discuss in Part 2A) means that soon you'll be able to give arbitrarily constructed commands to a computer for it to follow.
Five years ago you could use halting English to give commands like “Open word processor. File. Open. Essay dot doc. Enter. File. Print. Okay.”, but today it's possible to simply say “Open my essay and print it”, and in the near future combine a sophisticated sequence of commands in one sentence. “Use the Jones Project as a template for this new contract with the Smiths, figure out the budget for twice as many shipments, then have a printed copy ready for me along with a noon ticket to fly into Seattle.” It doesn't take any intelligence at all to follow those commands, it just takes a sophisticated enough set of rules to break it up into simple enough commands.
As for winning Turing tests and solving problems, AI research needs to unshackle itself from the language prejudice. A computer will not fool judges as long as it can only percieve of the world through grammar (check-out how bad the current state of the art is). They need to deal with the slightly less linguistic problem of representing and understanding the rules of the physical world.
The process of thought might be deceptively simple, like aerodynamic lift is, but without an equivalent to Bernoulli's principle we can't fathom how it might work. One obstacle is the learning process—the way a mind remembers what happened in the past, so it can apply it to decisions made in the present. It may be very separate from other functions of intelligence.
The complete series
While some scientists bang their heads on that problem, others have attempted to bypass it and get on with exploring the more useful application of reasoning, pattern recognition, and prediction. This brings us to the second category of AIs; the Hybrids. These are programs that, like all conversational AIs, have pre-digested memories supplied to them by humans, so they can get on with the far more interesting job of making deductions. And that's the topic of the next article in the series.
Last 20 responses and inbound links
(These are discovered in real-time and sorted by newest first. See how to get listed.)
- Not DisenchantedChris Wenham's LiveJournal
- USS Clueless
- Overclocker's Australia Forum - Semi-inteligent columnistNo, we're not the semi-intelligent columnist, but a reader is reminded of us in this forum thread. Registration required to read.
- Bluephod Infosystems
- Mark Paschal
