Slashdot Mirror
(Artificial) Mind Meld
Reader tewl points to this Wired article about a collaboration between the OpenMind project headed by Push Singh of MIT's Media Lab and Chris McKinstry's Mindpixel project. Neat to see these complementary projects getting along despite criticism each might have for the other. From the article: "The OpenMind and the Mindpixel projects will tie their databases together 'at the back end.' This means that any user data entered into either of the projects will be accessible by the other."
Here's a selection of data people have entered into mindpixel (rank as true or false, to validate them).
eat him?
okay, some of them are good, but they are all supposed to be context independent, and something that everyone would agree on. That means no opinion, no political campaigning, no paradoxes. If 10% of mindpixels database is complete garbage, of course it's never going to succeed. If it doesn't have an answer, people are just screwing up the system by entering it.
Right, what's important is the richness of representations. Daniel Dennett talks about an internal language of the human brain which he calls "mentalese", composed of representations that embody huge amounts of knowledge and link to one another in very rich, complex ways.
The meaning of a concept has mostly to do with the way it links to other concepts. Concepts link to the world outside the mind in two ways: the purely empirical way of sensors and actuators (my "apple" representation gets tickled when I see a red sweet edible object, or when I pick it up and bite into it) and the social convention of speech and writing. Speech/writing involves the least amount of actual meaning, so I'm not optimistic that these kinds of projects will get very far.
In another posting you mentioned the Cyc project. The interesting thing those folks did was to consciously plan an ontology, a roadmap of the ways that concepts could link to one another. This will allow them some freedom to deepen the level of understanding of which Cyc is capable. It would probably be good if the ontology could also learn from the data presented to it, rather than relying entirely on the conscious design decisions of its developers. They may not think of every important relationship between concepts.
Relating to rich representations, I came across another open-source program a couple of days ago called FramerD developed at the MIT Media Lab. It's a distributed database that's designed to handle millions of thickly interlinked records. The description says: One primary cause of brittleness, incompatability, and obsolesence in advanced applications is the premature codification of structures, protocols, and semantics. FramerD was designed to provide robust and efficient data management without extensive up-front specification of data and operations.
WWJD for a Klondike Bar?
I take issue with what it intends to (ever so vaguely) do. From the Website:
Eventually, it is hoped a GAC trained neural network will become indistinguishable from any human being when presented with any yes/no question/statement independent of whether or not GAC has seen that particular question/statement before
A Neural network is a Turing machine (a very large, hard to draw Turing machine ). This neural network will not solve the halting problem. Not too big deal, since I assume he meant "any reasonable statement", and exclude any problems that can be transformed into the halting problem. Still, it is an interesting point to bring up (I think).
Another issue: this neural network, can it reason about it's reasoning? Not terribly interesting if you can't get it to do that. Oh, it's still usefull if it can answery yes/no questions. You can always rip the network apart to figure out how it came to that conclusion. It's just very painful. And you end up with all sorts of numeric rules that are hard to give symbolic names to.
These are all just small sticking points. It would be interesting to see them addressed. I do have one large sticking point: it's a database. There is no intelligence. There is not intelligence in the facts. It's how the facts are used. And that's my problem with the entire project. It's quite clear that the intend to use a neural network trained on these facts. What kind of network? What sort of training? What sort of validation? Justify the use of a neural network over any other form of AI, suggest a new hybrid form. How will facts be encoded? Will a binary form of a grammar tree be presented? Input sizes to the tree will vary, how will missing data be handled? Will it only give yes/no values as output? Will it be a floating point number that we can assume to be a confidence value? These are very important questions that remain unanswered.
There seems to be no available information on how the facts will be used. Seems like a bit of a scam to create a database for resale to someone who might actually be doing some research.
McKinstry hasn't addressed why most real AI folks think his project is the equivalent of the Emperor's New Clothes: There will never be enough "mindpixels" to build anything constructive with, because the amount one would need to do so exceeds the number of atoms in the universe.
:-).
This makes no sense at all. The domain of all human knowledge is a finite domain. There is no indication at all that it would take more than the number of atoms in the universe to store it. In fact, we ourselves are pretty successful in storing it in just a fraction of the universe's atoms
I think you're confusing this with something like the chess game, of which has been proven that it would take more than the number of atoms in the universe to store all possible chess games in memory (thus finding the perfect chess game).
That said, I don't think the mindpixel project is going to be successful in any major way. This project is just a new trendy version of the classical attempt to model all human knowledge with symbolic information. A database of human knowledge modeled in logic just isn't flexible enough. Others have tried this and have failed. To name a 'few': Descartes, Leibiz, Hussertl, Heidegger, the early Wittgenstein, Winograd, Minsky, etc. etc. etc.. The list goes on and on, and these are not the least of names either...
This whole problem is often referred to as the 'frame problem' of AI (named after Minsky's concept of frames). This is not even close to being solved yet and in my opinion is one of the hardest problems ever to be encountered by science.
You don't need to be too bright too figure out what consciusness is - just objective. The trouble [for some people] is that the explanation treats the brain as a machine, and therefore:
;-), conscousness is based on connections that give access to *some* areas of our brain to others.
a) [in some peopls's view] devalues humans
b) reduces the role of religion to a mind meme panacea for the masses
c) allows that animals are also conscious
d) allows that machines could be made conscious
There's also the problem that the subjective experience of consciousness has a unique "quale" that seems not to be explained by any mechanistic explanation. However, the same thing could be said about the subjective experience of any sensory experience, just that consciousness is much more of a loaded issue. An explanation is always going to *seem *to leave something out because that's the nature of explanations - they reduce something mysterious to a colder set of facts.
I fully expect that a mechanistic explanation of consciousness will never be universally (or perhaps even widely accepted). Robots will be built that *will* be conscious, but many people will not accept that they are (how do I even know that *you* are, other than the reasonable assumption that since we're both human that you also have this "consciousness" thing I have myself?).
As for implementing consciousness, the problem is implementing the *rest* of an artificial mind that would support it. If you can build an artificial brain/mind that has most important human-level capabilities other than consciousness, then it would be easy to add the missing piece of the architecture.
Oh, what is consciousness: It's an inward looking "sensory" input. Just as our visual sense is based on the optic nerve carrying signals from external sensors into our brain (wah, but why does green look green?!
BizarroKiehl (and its stand-alone version, MegaHAL) is nothing like MindPixel and Open Mind. MP and OM attempt to learn facts and hopefully make connections between them; MegaHAL learns about where words are arranged in a sentence relative to each other, and makes no attempt to actually know facts.
:)
That said, I find MegaHAL more fun to talk to.
--
No more e-mail address game - see my user info. Time for revenge.
Win dain a lotica, en vai tu ri silota
These are both just open versions of The CYC Project. I have serious doubts about a project like this working, but if anyone *does* get it working, they'll end up doing it first. Unfortunately, it doesn't look like they're going to *release* anything to the public anytime soon.
However, I'd rather try to gather money to buy out/opensource cycorp than re-implement everything they've done in the past 16 years; they have a huge knowledge base already built, and a lot of code, and CYC can already do some interesting reasoning. (I know there isn't much there, but read what articles you can find; it's fascinating stuff)
And only using yes/no facts for data is just stupid; the computer needs to do some reasoning, and have some structure, otherwise, it would all just take too long! That's about as stupid as 'the table method' in AI. Even simple AI's can't necessarily be represented like that, so I hope there's more to it that I just missed.
...and for those people who think computers inherently will never be able to reason: go home; you aren't welcome here. I'll argue with your facts, but I won't cater to your prejudices.
---
pb Reply or e-mail; don't vaguely moderate.
pb Reply or e-mail; don't vaguely moderate.
I don't think we should mind if Mindspring mind-melds into this Mindgame. I have half a mind to to write Open Mind and Mindpixel, and give them a piece of my mind!
McKinstry hasn't addressed why most real AI folks think his project is the equivalent of the Emperor's New Clothes: There will never be enough "mindpixels" to build anything constructive with, because the amount one would need to do so exceeds the number of atoms in the universe.
I feel a little sorry for folks being snookered by this, because I think other more worthwhile projects are probably getting the shaft while college students spend their evenings typing in "Cocoa Puffs taste better with milk" and other variations on that theme.
Go ahead and look at the MindPixel site. See how vague it is about what exactly he expects to DO to all these bits o' "consensus fact" to transmogrify them into thought.
There's nothing there but mysticism.
I wish it weren't so, but it's so. This is Fool's Gold.
He who refuses to do arithmetic is doomed to talk nonsense.
Now we get to see what two different implementations of an AI network, with the same set of data, will do. Something of this scale has never been done before. We will get to see where each excels, and where each fails miserably, and hopefully some benefit will be gained by figuring out what parts of the different structural algorithms are best suited to AI. We may end up with one system that's orders of magnitude better at finding matches to loosely described data, but at a severe disadvantage on the turing test due to inability to mimic human emotions. Maybe one system will get the best of both worlds. Hopefully, each system will have at least a few advantages over the other, so something can be learned. After all, these systems are too complex to simulate or estimate mathematically. We can estimate the efficiency of an mp3 encoding algorithm, but neural nets are so vastly more complex, that beyond a certain point, you need trial and error. This case will give us trial and comparison, which is even better.
WARNING: there is a trojan on your
Similar to MindPixels, but far more entertaining, is BizarroKiehl, hooked in to AOL Instant Messenger. It learns from responses and replies based on what it learns, and is extremely amusing to boot.
From the old school, there's always AOLiza. She's not smart, she's not even that pretty, but she's the one all the guys want to talk to...
Kevin Fox
Kevin Fox
- Alice: Do you like books or TV better?
(I wasn't trying to fool it, I was trying to make it seem a little more noble that I watch TV a lot)Interiot: I find moving pictures informative.
Alice: Where are you going?
--
When MIT's Media Lab was founded in 1985 by Nicholas
Negroponte, the Lab emphasized computers and multimedia.
Ten years later it began its silly season with "Things that
Think" (chips in shoes or clothing that communicate with
the wearer, for example). But just then the Internet
materialized out of nowhere and caught the Lab with its
micropants down. Judging from its website, by now the MIT
Media Lab has made up for lost time by promoting projects
that expand e-commerce.
More interesting than anything the Lab has ever produced
is the fact that it's funded by big business. The Lab's annual
budget in 1995 was $25 million, mostly from 95 corporate
sponsors, half of which are overseas. While the Lab claims
that sponsors cannot dictate the research, it's also true
that grad students have to sign a nondisclosure agreement
before receiving aid, and sponsors often fund research that
is proprietary. Given this history, it's not surprising
that since the Internet arrived, the Lab has been chasing
the dot-com rainbow. But one has to ask: What about the
public sector? Where's the vision? Does anyone at the Media
Lab care?
This OpenMind project smells more like a rat than a mouse.
A computer knows only one thing, and it's the only thing
it is likely to ever know without insanely massive databases,
along with bloated fuzzy-logic programs that go by the name
of "artificial intelligence," but are really thinly-disguised
variants of brute force.
A computer knows this: one is not equal to zero.
Slashdot should try to stay clear of trendy hype backed by
big bucks. That includes Wired magazine, which received
start-up money from Nicholas Negroponte.
If it's only nonverbal, then you can't reason about it, you can't tell others your ideas about it in order to refine your concept of it, and we can't work towards putting consciousness in a machine (other than trial & error).
--
No one tried to explain the difference between an S and a 5, or even the difference between a letter and a number.
The end product was a best-in-class reader that could also reliably detect forgeries.
I don't think you could teach a person to read by showing them flash cards, having them guess the word, and then telling them whether they were right or not; it would take too long and the person would get very frustrated. Neural nets, on the other hand can learn this way.
These mind modeling projects differ only in scale, but the scale difference is gigantic. Can a neural net learn "common-sense" and natural language from a bunch of facts (some generally agreed upon, some not)? I think so, and am shoving stuff into GAC to do my bit.
An end result of a huge database of facts like "Water is wet" and "Picnics are fun" is not AI. Whether or not the back end can develop any kind of AI from this input remains to be seen. I hope so, because it would be really cool.
There are two kinds of sysadmins: paranoids and losers. I'm both kinds.
1. A neural network is not a turing machine. A neural network probably does not suffer from the halting problem.
2. Just because a neural network is run on a general purpose computer does not mean that it has the same qualities as the general purpose computer. There may be problems on which the computer would not halt, but simulating the neural network need not be one of them.
3. The halting problem doesn't have jack shit to do with artificial intelligence anyway. What's the problem? "We can't built something that's omniscient?" We're trying to build the equivalent of a human, and humans aren't omniscient. Similarly, Godel's theorems have nothing to do with AI either.
4. You can't "rip the network apart to figure out how it came to that conclusion". That's the whole point. If we could do that sort of thing, we could give up on all the AI research and just start building really fine-grain nMRI machines. But we can't, so we guess and check.
Regaring "it's a database, it's not intelligence", how do you know? What, you dream it? Divine revelation? Read it in a popular science book? "This blob of reddish gray stuff, it's just biological matter, operating according to the laws of physics, there's no intelligence."
I took a look at both of the projects: Open Mind associated text strings with pictures (discribing a picture, discribing a picture's contets, and so on), or one text string with another (explaining a fact, giving an example of a relation, explaining cause an effect, and so on). Mindpixel gets a collection of statements/questions in the form of text strings, and tries to get a consensus on whether the statement is true or false (or if the answer to the question is true or false).
But this seems to me to be the wrong way to go about it. While these projects will collect massive amounts of data, all that data is is associations between text strings. All they'll be able to do is detect that there's certain connections/correlations between certain words, and certain collections of words. This way of doing AI assumes that intelligence is just a bunch rules and mechanisms for manipulating symbols, with the symbols somehow representing chunks of information.
But what if you took these vast stores of information and replaced each word with word with some gibberish: "vut" replaces "car", "folp" replaces "clock", and so on. All the relations between words, and groups of words, remains exactly the same, but no human could understand it; all of the meaning would go out of it, because the meaning is being suplied from the outside, by the humans knowledge of what certain strings of letters mean.
However, if you were somehow to do the same scrambling to the vocabulary of a human's mind, so that this (formerly English speaking) human now used "vut" for "car" and "folp" for "clock", other people would eventually be able to understand and communicate with him; all of the meaning and information has stayed the same, it's just the labels that have changed. But for something like Open Mind or Mindpixel, the words aren't labeling anything; there's just relations between meaningless strings of characters.
The above argument is a (rather bad) summary of the argument that Douglas Hofstadter makes in the book Fluid Concepts and Creative Analogies. Anyone interested in AI should read this book. Douglas makes a very compelling argument that diving straight away into things like words and sentences is getting much to far ahead of ourselves, and that we first need to make tiny baby steps in AI before we can attempt to make an AI that really uses human languages.
Suppose you were an idiot. And suppose that you were a member of Congress. But I repeat myself.
Give a man a fire, and he'll be warm for a day, but set him on fire, and he'll be warm for the rest of his life.
Besides the vagueness, the idea of having the AI built solely to handle yes-or-no questions seems kind of limiting to me. I mean, just amassing an assload of facts is kind of trivial.
The user moderation system on Mindpixel is an interesting idea, but again, I don't think that it works very well either. There are too many cases where users either don't know the answer to the question that they're trying to moderate or they don't care enough to actually answer correctly.
The idea behind Forum 2000 - which I think we all know was fake - isn't a bad one: learn language usage from Usenet feeds, then dissect the messages in Usenet posts themselves. Has anyone done anything like that?
No one else knows what consciousness is either. Most of those who claim to know what it is explain it as "what people have that machines never will", gleefully ignoring the fact that humans are machines.
If anyone defines consciousness, then it's a goal you can shoot for. But Searle & Co. would never think of defining such a thing, because once you're pinned down, then you can be proven wrong.
I never suggested that reasoning is verbal. Just that one person probably can't figure out all of consciousness, so they have to communicate with other people. Also, it's often beneficial to write down ones thoughts and study them carefully... you'll often find incorrect assumptions that you made somewhere along the line.
--
Just because the computer outputs something physical doesn't make it any more intelligent. Now a computer that could learn to play a game well by reading a book on the subject -- that would be something.
You know, standard chat room conversations are probably even simpler than the original eliza. Forget the turing test, let's shoot for a program that can sucker a Disney executive into a meeting in Santa Monica -- not only would it require a pretty good pattern matcher, it could be self funding with blackmail funds.
For real fun, you can try the AI that won the last Turing test (to convince a human that it was another human, and not a computer) at www.alicebot.org.
- I don't care if they globalize against free speech. All my best free thoughts are done in my head.