pushblog

John McCarthy in town

2005-07-28T21:57:00.000-05:00

This is John McCarthy, me, Ed Fredkin, and Marvin Minsky at dinner. John McCarthy was in town for an interview about the Dartmouth conference, the event many regard to mark the birth of AI. It's rare that he and Marvin get together these days. I consider them the two people in the field most dedicated to building human-level AI.

Interior Grounding

2005-07-16T10:26:00.000-05:00

It has been hard to give this blog any attention while finishing my dissertation, but now that it is done -- I am now Dr. Singh! -- I will try to post a bit more often (although it is difficult to combat the Inverse Law of Usenet Bandwidth.)

I travelled last week to the AAAI meeting with Marvin Minsky, who gave the keynote talk on some new ideas he has been developing about how minds grow. The basic idea is called "interior grounding," and it is about how minds might develop certain simple ideas before they begin building articulate connections to the outside world. Marvin developed this idea partly in reaction to the popular desire to build AI "baby machines" that start with blank slates and develop ideas by extensive exposure to the outside world. The difficulty with baby machines is that the world is a confusingly complex place, and making sense of it is not simple matter. There may be ways to discover useful concepts in advance of exposure to the rich, outside world, and that -- when the mind is ready -- make learning about the outside world easier.

I think one reason Marvin likes the interior grounding idea is that it is very compatible with the Society of Mind theory, where the mind is seen not as a single agent, but as a society of simpler agents. In this society, some agents are directly connected to sensors and effectors that interface to the external world. But other agents don't see the outside world at all -- their concern is with the activities of other agents within the mind. Perhaps most of our agents are of this sort, concerned with problems not in the outside world but with other agents within the mind. If this is the case, then many of our mental agents may begin learning and developing, and develop sophisticated notions about how minds work, well before our sensorimotor agents have learned very much about the outside world.

big minds need to be SOMs

2005-02-27T21:02:00.000-05:00

Every now and then someone asks me why there haven't been many implementations of Marvin Minsky's Society of Mind theory of intelligence; instead, most AI people seem to use homogenous planners or inference engines and fairly uniform knowledge representation schemes, and these methods are often reasonably effective.

I've always puzzled about this myself, and I have some ideas for why this is the case. One possible reason is quite simple. It could be that the systems we've been building so far are simply too small to need the kinds of organizational principles that are described in the Society of Mind. Only now, when we are beginning to accumulate knowledge bases with millions or even billions of items, intricately represented and capable of being combined in many ways to produce many inferences, that society-like organizational principles are needed. Until now, using societal principles was overkill, like writing quicksort in object-oriented style -- procedural style is shorter, cleaner, and simpler. "Cellular" abstractions are beneficial mainly for larger programs.

Machine Desirantes

2005-02-21T21:06:00.000-05:00

Most television science fiction these days lacks advanced ideas -- generally, they seem to be run-of-the-mill soap operas or social commentary. But the other day I ran across an impressive animated show called Ghost in the Shell. The episode concerned the development of some self-reflection and community-reflection in a group of AI robots. I was reminded a little bit of when I first saw Max Headroom on television -- it's clear that this new show is ahead of its time, in the sense that it treats as commonsensical ideas like minds being able to enter and influence other minds, AI systems that share cognitive resources so that there is a blurring of identity, and there being a continuum of entities between people and AI systems. Max Headroom didn't make it past its second season, but the world it predicted is in many ways familiar today. I expect the same to hold true for Ghost in the Shell.

Dualism

2005-01-15T22:58:00.000-05:00

I ran across this peculiar thought experiment by David Chalmers.

He starts with the concept of the world as a simulation running in an external reality. But he looks not at the usual case where the brains of the agents in the simulation are also in the simulation, but instead at the different case where their brains are running outside the simulation and where the agents have no direct access to those brains. This is the situation, for example, of a typical "artificial life" system, or of a video game where the characters are driven by some AI system.

Chalmers seems to be claiming that, to the agents living in the simulation, assuming they've gotten interested in the question of what minds are and how they work, the most reasonable explanation is a dualistic one -- and that we ourselves should consider dualistic explanations as not so outlandish.

I see a couple of issues here.

First, it's not clear that the simulated agents need access to their physical brains in order to posit theories of how they work. People have been doing this for years by indirect psychological experiments. It may be possible for the agents to do similar experiments, ones that help them choose between different theories of their cognitive architectures. If they're smart enough and spend enough time on the problem, perhaps they would hit upon the right mechanism.

Second, Chalmer's scenario doesn't make me very much more comfortable about dualism. The analogue in the real world would be that the brain that we see is only a conduit to another "mental world" where the mind really happens. But why would the brain need hundreds of specialized centers to be, in effect, a glorified radio transmitter? One could make a case for this, but it strikes me as enormously unlikely.

Akihabara

2004-11-15T23:40:00.000-05:00

Just got back from a fund raising trip to Tokyo with Marvin and Ian. We were hosted by friends who took very good care of us, and I got to see quite a bit of the Tokyo area. It was great fun to visit Japan again -- the last time I was there was just about 10 years ago. My favorite part was wandering the electronic components stores in Akihabara. I didn't realize Japan had such a vibrant hardware hacker culture. There were several multistory buildings packed with dozens of small shops that sold parts of all sorts. I haven't seen these sorts of shops in the Boston area (with the exception of the monthly MIT Swapfest) -- presumably because mail order is so efficient in this country. But I remember how nice it was as a kid in Montreal to be able to just run over to a well-stocked electronics store when I needed a part. One question I had was how the Japanese proprietors could afford to run such small businesses -- as in the one in this picture.

the meaning of knowledge

2004-10-25T21:32:00.000-05:00

I met with some officials from the Department of Defense's National Security Agency today. I was describing our work on commonsense reasoning when the question came up of what did we mean by the word "knowledge." In our systems we usually start off with a corpus of commonsense facts, stories, descriptions, etc. expressed in english, but which are then converted through a variety of processing techniques into more usable knowledge representations such as semantic networks and probabilistic graphical models. My suggestion was that, from the perspective of the computer, only the latter forms should be considered "knowledge" because they could be put to use by an automated inference procedure. But in the long run, as our parsing and reasoning tools get more sophisticated, we may come to be able to use more and more of the collected corpus.

Generally, the word "knowledge" is very inclusive because the AI community has discovered a vast array of knowledge representation forms, and every one of them is useful for some purpose or the other. Thus, the more important questions may not be what is and isn't knowledge, but given some knowledge, questions such as the following:

For what purposes can it be used? When is it applicable? Is it true? According to who? Under what circumstances? Who might find this knowledge useful? Is it expressed clearly enough? Are there other units of knowledge that may be useful in conjunction with this one? How long should we expect this knowledge to stay relevant? How might have this knowledge been acquired, and from where might we acquire more like it? What background might you need to make sense of it?

And so forth. The point, I suppose, is that like most words that point to complex ideas, understanding the word "knowledge" requires that we consider its many contexts of use, and the issues that show up in those contexts.

Why No Vision?

2004-10-24T00:14:00.000-05:00

Why is it that computer vision has proven to be such a difficult problem? The strange thing is that computer graphics, which one might regard as the inverse problem, is rapidly closing in on achieving photorealistic rendering of scenes. I'm also puzzled because recognition problems are typically simpler than generation problems. It's certainly true that computer graphics has benefited from much commercial development and Moore's law, but faster computers should help recognition tasks as well.

One idea is that vision suffers from the same kind of problem as does commonsense reasoning, namely, the lack of large scale knowledge bases about the kinds of objects and materials in the world, what they look like from different angles and under different lighting, and so forth. But if this is the case, and computer graphics has advanced so far, it should not be difficult to generate a suitable such corpus with a moderate investment -- a corpus of images, ground truths in terms of 3d and other types of surface models, and connections to more general commonsense knowledge.

the paradox of Wikipedia and Open Mind

2004-10-20T20:20:00.000-05:00

I ran across this quote from Larry Sanger, one of the co-founders of Wikipedia, about Wikipedia:

It must be full of a bunch of crank submissions, vandalism, and plain old sophomoric stupidity. But it's not. It's not half bad. In places, and increasingly, it's of very high quality. And that's even more paradoxical.

Our Open Mind Common Sense project was greeted with similar skepticism by many in the AI community. I remember a meeting at IBM two years back where I described how we were trying to build a tool that would let members of the general public collaborate to build a commonsense knowledge base. I should have brought some asbestos underwear! The audience, largely members of the mainstream knowledge representation and logical AI community, was livid at the possibility that we might be able to use facts expressed in natural language contributed by the untrained masses. Some of their concerns were valid, but I was astounded at the level of conservatism they demonstrated. I'm happy to say that since then we've had quite a bit of success using the data collected by our Open Mind project.

It's great to see the success of Wikipedia, which confirms that it's possible to create a useful knowledge base from the contributions of a great many people, and especially, that it's possible to manage the vandalism and to some extent the disagreements. I now wish we had been even more courageous with Open Mind, by allowing people to edit and repair each other's contributions. I was worried about vandalism, but perhaps the problem would never have been as severe as we feared, and perhaps we would have developed strategies to deal with such problems.

commonsense and the practice of AI

2004-10-12T17:51:00.000-05:00

I had a chance today to spend a little time with a division director of the National Science Foundation. I showed him some of the work we are doing at the Media Lab to give computers more "common sense." I spent much of our time on the commonsense knowledge bases and reasoning tools we are developing, and a little of it on the cognitive architecture we are developing to coordinate the use of these systems.

What we are doing is considered very unconventional, but I hope that the larger AI community will soon come to see the value of the commonsense approach. My belief is that the practice of AI will change dramatically as large-scale commonsense resources become more commonly available. Instead of machine learning from scratch, much machine learning will be concerned with learning in the context of substantial existing knowledge. Instead of hand-coding small knowledge bases for their domain, researchers will augment existing commonsense knowledge bases with the additional knowledge their domain needs. Instead of perceptual systems operating mainly bottom-up from the signal, they will come to use large amounts of top-down knowledge about a wide range of objects, situations, and events, and so forth.

Fresh Ideas

2004-09-24T09:21:00.000-05:00

I swung by Marvin's place yesterday and he showed me a section he is drafting for his upcoming book The Emotion Machine. The basic idea is that there may be sections of the brain that are "consciousness activity detectors" -- agents that recognize that self-reflective processes are active. This could explain why there seems to be some uniformity to the phenomena we call consciousness, even if there are really many different kinds of processes at play.

It's awesome to see these ideas fresh from his mind; generally, it's been fascinating seeing the book evolve over the past ten years. It's very much been like watching someone else's child grow up -- you see the incubation of the early, ill-formed notions, the trying out of different variations, the changes of interest and focus, the sudden spurts of development, and eventually the maturation into a distinguished member of society. I expect this book will be one for the ages.

older, better?

2004-09-20T00:00:00.000-05:00

Is it a myth that people slow down as they get older? In my experience, complex ideas now seem simpler, I'm better at solving problems, and generally I'm faster at learning new things. I'm certainly more critical of my ideas, but my sense is that the ones that get through my filters are better than the ones I would produce a few years ago.

Someone once told me about a class they took with Gerry Sussman, one of the great teachers here at MIT. One day he came in and started writing equations on the board. As he went on the equations got more and more painful to follow, and eventually one of the students gave up and asked him, how can you possibly understand this complex stuff? Gerry's response was, "Well, I couldn't when I was your age -- but when I turned 26 I grew three new registers."

the term "symbolic AI"

2004-09-19T12:26:00.000-05:00

There seems to be a basic misunderstanding among many amateurs and students of AI (and shamefully, among some professionals as well) about what "symbolic AI" refers to.

Today there are many AI techniques that employ symbols -- logical AI obviously, AI based on the use of frames and semantic networks (which is distinct from logical AI because they often do not have a clear logical semantics), genetic programming (where one searches through spaces of symbolically represented genotypes), and Bayesian networks (where the nodes are often labeled with meaningful symbols.) Neural networks often do not have anything you could reasonably call a symbol in them, but there are communities within AI that have developed many ideas about how to express structures such as frames and semantic networks in connectionist substrates.

Yet despite this variety I often run into students who lump logical AI, frames, and semantic networks into one category; genetic programming into another; and Bayesian networks and other probabilistic models into a third. There seems to be no special reason for this lumping other than the so-called symbolic vs connectionist debates of the late 80s and early 90s which established this false dichotomy.

If AI is going to become a mature field we need to start teaching students to make finer distinctions -- which means that they will need to be trained more deeply in more subject areas, including all of the above.

Commonsense Reasoning Wikipedia entry

2004-09-18T10:32:00.000-05:00

I started a Wikipedia entry on Commonsense reasoning. I started it with the line:

"Commonsense reasoning is the branch of Artificial intelligence concerned with replicating human thinking."

I'm not sure everyone would agree with this definition--some might argue that it is the branch of AI concerned with particular techniques such as default reasoning with logic, in which case it is applicable to a wide range of problems beyond those people solve.

Or, one might choose to regard Commonsense Reasoning as concerned with solving the kinds of problems people solve--e.g. understanding sentences, recognizing scenes, interacting socially, etc.--but not necessarily in the same ways people solve them. I tend to fall more into this camp.

LifeNet coming soon

2004-09-17T22:20:00.000-05:00

We're getting closer to our release of LifeNet, a commonsense reasoning system that uses a 1.5 slice DBN-like probabilistic model (it's actually a pairwise Markov random field.) This isn't the best way to represent commonsense knowledge, but I suspect it's not a bad substrate on top of which to layer other techniques, plus we have a number of ideas about how to rapidly grow this model.

Marvin Minsky once drew the following diagram as a way to think about how to relate the various representations discussed in The Society of Mind. LifeNet was originally intended to represent knowledge at the Transframe level, and it does indeed have some of the properties of the knowledge at that level. But it is probably a better fit to the Microneme level, with the extension that LifeNet associates propositions across time slices as well as within time slices.

a practical problem in modern computational epistemology

2004-09-10T20:24:00.000-05:00

To me the idea that humans represent knowledge as stories is one of the more compelling arguments for the need to have rich knowledge representations like those used by Cyc. Simpler representations simply lack the power to express the variety of content and structures that stories contain -- situations, places, times, objects, events, beliefs, goals, emotions, and so forth.

That said, I am dubious of the proliferation of symbol names that seems to be the product of most ontological approaches to AI. The problem is that it is difficult to decide when to stop. Pat Hayes once recounted to me a story about a group of ontological engineers arguing about whether a painting that was hanging on the wall was “in" the room--but then an even fiercer argument broke out about whether the paint on the wall was "in" the room! It's certainly an interesting exercise to produce and refine such distinctions, but my sense is that this is not a well-defined task outside of some purpose or goal that guides the production of these distinctions.

Stories, however, open up an interesting new possibility. Rather than using a large collection of special symbol names that distinguish between an ever-increasing collection of cases of "in-ness", we can instead say ‘in’ as in the story STORY-532. At some point symbolic distinctions could begin to be made by exploiting their extrinsic contexts of use to provide a basis for further refinement, as opposed to trying to shoehorn that context into the symbol name itself.