I think there are two problems with your strategy. First, I think that it would be a shame to split up AI into a lot of small sub-disciplines if we can avoid it. I think that the various communities within AI have a lot to learn from one another. These communities could be mutually supporting, as long as we can maintain some level of mutual respect and get over the idea that not-so-formal work on human-like AI is somehow confused and second-rate. Many of the specialists in super-human kinds of AI still dream about solving the larger problem, but they are frustrated and have turned to sub-areas of AI that offer more immediate practical results.

Second, those of us working on human-like AI are not in a position to throw out all the super-human specialists, even if we wanted to. They are a large majority of the field now, and have been for some time. So all we could really do is secede. Some have done this, renaming their new group “Artificial General Intelligence” or AGI. It’s unclear whether this splinter movement will thrive. (It’s interesting to me that a number of papers at recent AGI conferences talk about the need for a new mathematical/theoretical foundation to enable forward progress in AGI. To me, that’s the sort of thinking that got us into the current situation, but I wish them well.)

There is also a lot of work on “biologically inspired” AI. That’s not quite the tack I would take — I think it’s worthwhile to study human-like general AI without necessarily focusing on brain modeling — but it is one respectable way to rule out narrow work on super-human topics. There have been a couple of successful workshops (not narrowly focused on brain modeling, despite the label) at recent AAAI Fall Symposia.

– Scott

I don’t see how this could be a problem, as long as these narrowly-defined, optimal or near-optimal technologies are spun off to domain experts ASAP. Let optimal planning papers be published in operations research journals, papers about chess/poker-playing programs in computational game theory journals, and papers about statistical inference in statistics journals.

Good questions. Here are a couple of quick responses.

Regarding your first question: While my work on Scone is focused on symbolic representation and reasoning, I don’t believe that this is the whole story of human knowledge. The symbolic part is a very important component of human cognition (perhaps less so for animals), and it is the component that connects most directly to our linguistic mechanisms. In fact, I often speak of Scone as a representation for “all the things we can easily describe in English”.

In addition to the symbolic representation and reasoning part, I think we humans also make extensive use of visualization – actually thinking in terms of images, probably using 2-D buffers and processing machinery borrowed from our visual system. For example, if I ask you what is the best sea route from London to Venice, you will probably have a strong subjective impression of looking at a map and tracing a route – perhaps a mentaldotted line – from London to Gibraltar, south around Sicily and the boot of Italy, and up the Adriatic. (Your more symbolic machinery may butt in to advise you to avoid the Strait of Messina, with its famous vortices and rocky hazards.) Subjective impressions can be misleading, but I think this one corresponds pretty closely to what the brain is actually doing.

In addition to 2-D imagery, I think we have specialized mechanisms for representing 3-D imagery, low-level motion planning, sound sequences (as when you remember a song), and probably a few other things. I’ll have more to say about these additional representational mechaisms in an upcoming article for this blog.

An interesting question is whether these different knowledge mechanisms are so intimately connected that it makes no sense to work on the symbolic part separately, and to hope that we can connect this to the other representation/reasoning mechanisms later. I think we can get away with this – probably, more or less – and progress certainly will be much faster if we can “divide and conquer”.

Regarding your second question, I should say that I’ve read a little bit of Dan Sperber. Much of what he says seems sensible to me, but most of it is vague enough that it can be both right and wrong at the same time. Or maybe I just haven’t properly understood what he is trying to say – perhaps some ideas of my own have inoculated me against the epidemiological spread of his cultural idea. :-)

I don’t think that Scone imposes any particular context or cultural viewpoint on what can be represented. In fact, using Scone’s multiple-context mechanism (the subject of several more articles, coming soon), we can represent multiple, possibly inconsistent explanations or viewpoints of the same object or episode. One context might explain a sequence of events as the result of conventional physics and causality; another might describe the same events in terms of magic or meddlesome kami or flows of phlogiston. You can sit in either context and reason about the forces at play and about what might happen next.

In fact, some colleagues and I are trying to get funding to explore how we can use Scone’s multiple contexts to organize knowledge about various human cultures, and how these cultural ideas and viewpoints interact with everyday actions and language.

– Scott

Second, as an erstwhile linguist, I am concerned about the possibility of concepts existing outside context (viz. Dan Sperber on relevance). Even an elephant, of which most of us can generate a mental picture if we have seen 2-D representations from childhood, as a concept has such dynamic attributes as functional load, a set of frequencies and contexts that are both language- and culture-bound. Certainly, such probabilistic attributes can be ignored as accidents attached to some kernel notion or accounted for by reference to some other features of a given discursive event, but this kind of thinking is already metaphorical, hence bound to the symbolic plane. And, hence, incapable of allowing us to determine whether or not non-symbolic associations might somehow represent the processes we seek to emulate more accurately or usefully. While you dismiss the possibility of “magic” (in an earlier post) as a pathological case to explain the workings of the mind, there do seem to be actual qualities that come into existence merely by virtue of accidental collocation or provisional context, that would be difficult to include in some a priori and context-free ontology. At the anthropological level, in fact, the idea of magic seems to represent the human recognition of the failures and lapses of symbolic systems to account for all of cognition and perception. Cause and effect, which seems to be a notion that is a consequence of language operating in real time, is itself bound to the symbolic plane. What may be desired is a representation that does not avoid cause and effect, of course, but also permits some calculable value to be assigned to synchronicity, to the provisional. The “qualitative.” Thus, again, from a processing POV, I like Sperber’s sense of the “low cost” of relevance, and such a nervous signal should somehow be attachable to any conceptual ontology.

Sorry for the delay in responding to your queries. Busy times…

>> In your KB, how would it be possible to replicate the phenomenon that humans think examples of a concept are often placed on a continuum, where some examples are more archtypical than others? (I.e. robins vs ostriches as an example of the concept “bird”.)

One simple metric would be to count the number of exceptions used in describing an ostrich or whatever. A more sophisticated metric would look at each of these exceptions and would see how unusual it is in the population of subtypes and instances. For example, birds fly by default, but we have many instances of flightless birds. On the other hand, we have no instances of modern birds with teeth, so that would be more atypical. And we have no examples at all of living animals made of steel, so that is still more atypical.

>> How are you going to deal with context and environment for a concept? For example, a lamp on a desk is common. A lamp taped to a horse’s head isn’t.

We could represent context by a simple “usually-found-in” relation, or via the context mechanism in Scone. I won’t go into details on that right now.

When doing recognition, context-based expectations enter into the matching process in a way that is very similar to observed features. I might be trying to recognize a large, animal, in Africa, that is very big, gray, with a long cylindrical nose… The “in Africa” part acts very much like the other features.

Often context and expectation can trade off against features. If I walk into my office and it’s dark, I will assume that the big dark blob near the wall is my desk, without having to see it clearly. My expectation is driving the recognition here, and it requires very little sensory input to confirm this strong expectation. On the other hand, I see a two-meter-long alligator in my office (some of my grad students have an odd sense of humor…), it may take me a second or two to recognize the thing because it violates all my expectations. But if I see it clearly enough, the sensory features will overwhelm my expectations and I will probably start running.

>> The KB you describe appears to be based on relationships that are cast in stone. I.e. An elephant is a vertebrate. … Wouldn’t it be better to have a KB where relationships and hierarchies are more fluid, learned and changeable instead of fixed forever and entered laboriously?

You’re right. Scone’s categories are not cast in stone. It is easy to add new types (each with a description of the typical instance). Sometimes these are handed to us – as you say, we learn the concept of “vertebrate” in school or form a book – but others are created by learning. – that is, by noticing that a bunch of individuals have some features in common and creating a type to contain these. If you have never heard of an elephant, but moved to Africa and saw all these big, gray, long-nosed animals walking around, you just create a new (informal) type to hold these things, and you move their common properties up to the type-description. Then, for each individual elephant, you can add any specific features and maybe cancel some features that would otherwise be inherited.

You might worry that creating a new class is a dangerous move, because it might conflict with some other (perhaps better) categorization that we learn later. But a multiple-inheritance system like Scone is very forgiving: creating one set of categories doesn’t preclude creating a different set later. And if an elephant inherits “has a heart” from several superior classes, that redundancy doesn’t create a problem unless the heart descriptions are incompatible.

Cheers,
Scott

How are you going to deal with context and environment for a concept? For example, a lamp on a desk is common. A lamp taped to a horse’s head isn’t.

The KB you describe appears to be based on relationships that are cast in stone. I.e. An elephant is a vertebrate. Wouldn’t it be better to build a KB that can develop its own hierarchies from examples? For example: a child would probably know that an elephant has a heart but an ant doesn’t. This knowledge probably wouldn’t be based on anything sophisticated like “vertebrate”. Rather it would be based on something like size and similarity from examples. So a child might learn that people, dogs, cats and horses have hearts. From that, the child would probably extrapolate that elephants have hearts. Ants are too small to have hearts. It’s only later in school that such sophisticated concepts as “vertebrate” are learned. Wouldn’t it be better to have a KB where relationships and hierarchies are more fluid, learned and changeable instead of fixed forever and entered laboriously?