Overview of AGI Research
Posted by Jeriaska on April 29th, 2008
Photo courtesy of Hugo de Garis
The first panel discussion of AGI-08 was on the subject of research methods for artificial general intelligence. Session chair Eric Baum started off by responding to presentations by panelists Jonathan Connell, Joscha Bach, Wlodzislaw Duch, and Pei Wang. Questions on the first of the conference’s presentations then were taken from the audience.
The following transcript of the Overview of AGI panel discussion has not been approved by the speakers. Video is also available.
Overview of AGI Research
Eric Baum: I will make one comment that will take us back to the very beginning. In the first talk, [Seven Principles of Synthetic Intelligence], you included a skeptical quote of Leibniz: “mind as the machine.” I am pretty confident he was reacting to the earlier work of Spinoza, who had deeply offended him by realizing, in a way that I think is incredibly perceptive given what was around at that time, that the brain is a complicated machine and that consciousness and everything that goes along with it is just a description of what is going on in that machine. Nothing more than that. It is amazing to me that Spinoza was able to figure this out back when he did.
I would react also to the second talk [Four Paths to AI] in one other sense. In your core values, I would put economic modeling. More generally, I would say one thing that is needed is some principls that make the thing be grounded and generalized. Not to get back to Occam’s razor, but adding lots of stuff on top of it would not necessarily make it generalize. Somebody has to constrain it, and you would like to have some principle that makes things constrained. An economy might do that, so that it will generalize and do the right thing in new circumstances.
Joscha Bach: You addressed this point that we want to have something which is grounded and general at the same time. I think this is one of the very crucial things. I try to make this point as a benchmark. I think Wlodzislaw did the same thing when he came up with the twenty questions test. For me this is a very interesting point. I think that apes will fail the twenty questions test, and so will most children. In fact, a lot of adults may fail the twenty questions test because they are not intelligent enough in order to succeed.
The question for us is, should we build apes and extend them a little so they can manage to survive the twenty questions test? Of course, we are not necessarily looking at something that has to be human in order to be intelligent, but we are looking for something which has to encode somehow the same kind of constraints of representation and information processing. This is the point where we need to start. Maybe, to be like a human is the best way to embody the same kind of constraints of information processing and representation. The crucial point here is not that we need to be grounded in a physical reality, but in some deep sense it ought to embody these constraints.
This is maybe an engineering question–a question of the methodology that we need to choose. I think this is what I want to learn more when I come to this conference. I have this opportunity to be here with all these nice mad scientists which pursue this goal that nobody else wants to aim for.
Wlodzislaw Duch: Let me ask, what is the reaction everybody has? Is embodiment not simple use and practice? Aaron Sloman has recently written about embodiment. Embodiment for language, for me, is the use of certain concepts in different contexts. When I think of temporal concepts, that is different, because that relates directly into meaning in the world. When I think of certain mathematics, should there be any embodiment in this? That is probably doubtful.
Ben Goertzel: A brief comment on that–if you look at Lakoff and Nunez’s book Where Mathematics Comes From, they propose precisely the argument that mathematical intuition is grounded in physical interactions.
Wlodzislaw Duch: I know. A lot of engineers have not read Lakoff, and their intuition is not so grounded. They see complex numbers and they do not think about self-referential regulated processes. I’m not sure how much embodiment is really needed in this case.
Sam Adams: I would like to respond to that a little bit. When you talk about what kind of embodiment, a lot of people who focus on embodiment like I do, we measure that in terms of how many sensors, how many actuators, how connected are you to environment? I think it is really similar to a lot of what your talk said. In all those examples, there was always one bullet in there somewhere: no large-scale experiment. There is a feeling generally across all these areas that there is some sort of a tipping point where before you have enough, you don’t get interesting results. Once you go past that number, you start getting different results.
A lot of the algorithms that people have done in AI over the years work in the small, but do not work in the large. A lot of theories may very well work in the large, but cannot work in the small. I think when you say, “How much embodiment,” it is part of a more general question about what is the scale of the system you are working in. We are working on far too small scale systems right now.
Wlodzislaw Duch: That also very much depends on the application. If you are talking about sensors, you need robots. The relation of perception to my actions is important because we can focus on things we want to really see. There is a fellow at Stanford who has done this in vision, which is very helpful. He says that from a distance the robot does not have resolution to see the object, so it cannot recognize the object. But when you focus on just a single object, then you see the object much better, and the whole biological systems seems to be very efficient in this way. In many other domains, the application of AI as a personal assistant, this type of embodiment can be really quite secondary.
Joscha Bach: Can we have embodiment in mathematics too? If we look at physical reality, it is like a small dirty pond, and from time to time we catch something and encode it. When we do mathematics, something similar happens, because mathematics has its own constraints. It is not a small, dirty pond but a big, crystal clear ocean. We wave our small sieve of reason around it and look for similarities, then we encode them.
All the ideas that say that when we do mathematics we abstract over our experiences, maybe this is true for humans and for the trajectory that most humans approach mathematics at first, but when they are mathematicians, maybe they are different. Maybe an AI could be embodied over mathematics, but it is pretty likely that you could not talk with that anymore, as many people cannot talk with mathematicians, and vice versa. I think it is conceivable that mathematics itself is an environment where rich embodiment is possible.
Audience: This is more of a general comment–it’s not directly related to emdobiment–but it seems the general state of AI research currently is rather like the research on life a hundred and fifty years ago. A hundred and fifty years ago we kind of new what makes things living, but our ideas of what life was were pretty naive. Rather than being dualists, we were vitalists. Then we came up with evolution, discovered the structure of DNA, and now we have biology.
Despite having a pretty good grasp of biochemistry, embryology, and a pretty good ability to see and actually manipulate the stuff of life at every step, after a lot of research we might be able to soon synthesize a single highly idealized cell. We used to study life the same way we are now studying AI or cognition. One hundred and fifty years ago we took our folk ideas of what it meant for something to be living and hoped that if we jury rigged them all together that some kind of life instantiating Rube Goldberg machine, we would create something that moves on its own. It seems the key problem in AI is that we are not even sure what it is that we are talking about. Life life researchers of one hundred and fifty years ago, we have all these folk ideas of properties that we want for artificial intelligence but we do not have any idea of what is going on as a cognitive equivalent of the cell level.
Life is fundamentally unintuitive to us, and requires that we look at it at its lowest and most general levels. Still, a lot of people do not understand how life works today. So I guess my question is, why do we think that the structure and function underlying cognition and artificial intelligence will be more intuitive and well described by our folk ideas than the structures and mechanisms that ended up underlying life? In other words, why are we more interested in synthesizing at the highest levels something we do not understand instead of trying to figure out what it means to represent it at its lowest level?
Wlodzislaw Duch: I think we are very much trying to understand. Some of my slides, which I did not have time to show, were on what we know from neuroscience now about how these things are implemented. I think we know almost everything, but we do not have much experience with building very complex systems, and we have not done even basic things.
When you think about language, you look at what is in there–there is some ontology, but I cannot find a common sense ontology. Ontologies are very specialized. If you are a biologist, you find beautiful ontology that tells you all the details, but that is not what normal people have in mind. We have not yet really created resources that will, for instance, describe every concept in relation to others. People do small semantic networks, and they work in their small domains, but if we had something on the subsymbolic level and could use it in different domains, we could really make some progress. That has not been done.
David Friedlander: I would like to draw an analogy between embodiment and dissipative systems in physics. These systems take energy from the environment and they undergo spontaneous decreases in entropy and self-organization based changes. I think an intelligent system needs to take information from the environment in order to learn and adapt. That is the kind of embodiment that matters. It does not necessarily have to be in the physical world, but it has to have an information-rich environment and condense the incoming information.
