Future Current

Understanding why the original project of Artificial Intelligence is widely regarded as a failure and has been abandoned even by most of contemporary AI research itself may prove crucial to achieving synthetic intelligence. Here, Joscha Bach of the Institute for Cognitive Science, University of Osnabrück, Germany took a brief look at some principles that we might consider to be lessons from the past five decades of AI. The author’s own AI architecture, MicroPsi attempts to contribute to that discussion.

The following transcript of Joscha Bach’s AGI-08 presentation “Seven Principles of Synthetic Intelligence” has been corrected and approved by the speaker. Audio and video are also available.

Seven Principles of Synthetic Intelligence

Coming from a background of cognitive science, with a particular focus on computer science and philosophy, I am mostly interested in looking at cognitive architectures as a tool of understanding intelligence. Cognitive architectures can act as testable theories on how the mind works, and what constitutes general intelligence.

When we do AGI, we are trying to understand the mind as a machine, an idea already stated by Gottfried Wilhelm Leibniz. To paraphrase Leibniz: If we want to look at the mind scientifically, as an engineer or as a natural scientist, we capture it as something that brings forth thoughts, experience, and perception. If we enlarge this something, if we can see it as something like a huge mill, we can see how the individual parts work on each other. This is what we are trying to do when we try to understand the mind by constructing a working model of it, the same way that physicists try to build a formal theory of the universe and see if it works out in their simulations.

AI scepticism

Sadly, this is not the whole story, because Leibniz himself did not believe in that proposal. The full quote is: perception, and what depends on it–that is, cognition, thinking, the mind–cannot be explained, because otherwise it would be a machine. It is not conceivable that you would have parts working upon each other to bring forth cognition. Leibniz did say this many centuries ago, and Leibniz conviction is far from being dead.

Roger Penrose, for instance, says the same thing when he reminds us that understanding and perception cannot be simulated computationally. It has to be something which goes beyond computation, beyond formal theories, and he thinks it may be quantum mechanics. For some reason he believes that quantum mechanics itself goes beyond what we can capture mathematically, so it is a mystery. Since consciousness, the mind and so on are mysteries themselves, hey!–maybe it’s all the same thing.

John Searle also goes in a similar direction when he points out that computers can only do syntax–they only can do symbol manipulation–and understanding is something that intrinsically has to go beyond symbol manipulation. It cannot come from mere symbol manipulation; it has to come from something different–maybe from the intrinsic properties which only are inherent to biological humans.

To put this more lyrically, the experience of the human being is not transferable into a machine. It is something that computers cannot have; computers cannot be creative. There are many more angles to the same argument, many variations to the same tune: We have a cultural consensus in our Western society, more or less, that computers cannot and should not be intelligent–maybe they cannot because they should not. This “Winter of AI” that we are witnessing is far from over.

We are facing very strong cultural opposition. When we are pursuing general artificial intelligence as a subject, we find ourselves in a similar position to somebody who does genetics, and tries to build a genome to create a cell or organism, and the funding agency is populated by creationists. Even though Europe is not into creationism, the AI scepticism is even worse in Europe than it is in the US.

Troubles in AI research

But is cultural AI scepticism the only problem? Of course not. Artificial intelligence research has discovered a lot of traps that it managed to fall into on its own. We are usually not really discussing what we are up to, what we are really trying to build. What kind of modules, techniques, mechanisms we want to implement. We have this general direction, artificial intelligence, and then bend it many ways to make it fit certain tasks. We tend to diverge from the original question, and we tend not to formulate this question in a very precise manner.

Avoid methodologism, and stay away from methods that do not lead to strong AI

On the other hand, we fall into methodologism. That is, we develop a set of methods, and then we divide the intricacies of these methods into communities and sub-communities. We specialize and we publish in these communities, and eventually these methods do not add up again. We do not get unified architectures; we do not get whole pictures. Rather, we get individual modules.

The mechanisms that we build tend to be ungrounded. They tend very often to be symbolic–they do not scale up. We have too many of these approaches in AI.

On the other hand, we have too many of the dumb robotic approaches, too. We have lots and lots of interesting work with robots, which has a lot of value on its own–just as the symbolic processing has value on its own–but maybe not with respect to getting to the goal of general intelligence. Typically, for instance, there is the lack of integration of motivation and representational structures in AI.

Most of all, AI suffers not only from a lack of funding but from a lack of conviction. Maybe these go hand in hand.

What can we learn from that? First of all, we should aim at whole, functionalist architectures.

The functionalist requirement

What do I mean by functionalism? What you can see here is something like an fMRI image of a combustion engine. Actually: It is not an fMRI, but an infrared image, but there is a big similarity, which does not stop with the colors. If we look at this image in motion, we see lots and lots of very interesting things. We can see if there are defects in the machine. If the machine does not run the way it should, we will find certain correlations which are very typical in this image. But the bad thing about this image is that it does not give an explanation of how it works, because it does not identify the functional parts. What we do in cognitive science research these days, especially in neuroscience, is that we mistake such an image with an explanation of how it works.

What we want to have is something like this: a functional schema of the mechanism. Of course there is a lot of value to an fMRI image, but only in the context of a functional explanation. We need to have a functional explanation to impose it upon the descriptive fMRI picture, and see what these individual parts actually mean. In order to get there, we need to have a conceptual decomposition of the whole thing into its parts and how they interact with each other.

Aiming methodologically for the big picture

Second, let this question of how to achieve general intelligence define the method, and not vice versa. It is not going to help us to have all these methods and then try to accumulate them or reify the different areas of AI. If you go to an AI conference these days, you will have papers on description logics, on agent technologies, on robotics, game theory, semantic web, and many, many different areas. It is not as if someday we are going to merge all these things into artificial general intelligence–it’s not going to happen. What we need is to define our methodology according to our question. It is pretty much the same thing as Ben Goertzel just said. We should aim for the big picture, not for narrow solutions.

Build grounded systems

We need to have a conceptual decomposition. We should have an understanding of the individual parts which we have to integrate in such a system, what the theory of mind is with respect to other agents, how memory works, how perception and action are integrated, and so on. We should also aim for grounded systems. That is, systems which do interact with their environment. On the other hand, we should not get entangled in the philosophical symbol grounding problem. There is nothing mystical in reality which you have to touch in order to gain intelligence. We should also remember that many, many embodied systems which we find on our planet, the vast majority of them are not capable of general intelligence.

Do not wait for the rapture of robotic embodiment

We need to have representations which are adequate to represent the world and to tackle the combinatorial explosion that comes with that. We need to have representations which relate to how we interact with the world, not just abstract symbolic descriptions. We should not wait for robotic embodiment in order to get there. Robotic embodiment is very, very costly–not just because of the cost of robots, because the price has gone down considerably–but in order to get the robot to do certain things, we are going to spend a lot of time and focus.

I have worked on robotic soccer for a few years, myself. I have learned that if we do this in a very closed environment with a limited number of representations which we encounter, it is not as if we just arrived there and conquered the field the same way that chess was solved. It is a very interesting task in its own right, because we have a discrete environment with many unforeseen features, and so on. Eventually, this is probably not going to end up in general intelligence, but in some kind of secondary solution that is very good at playing robotic soccer. The environment which you have there is considerably less complex than many virtual environments I have been playing with.

Choose a problem that has general intelligence as a prerequisite

The big challenge which you have here is to find benchmark problems for AI, which require intelligence in order to solve. Looking for benchmark problems apparently is a very difficult task, but a necessary one. We should also aim for autonomous systems. Intelligence is not something which is related to problem-solving. It is about finding the problems in the first place. Intelligence is not the answer to some resource allocation problem–it is a very complex control task. We are organisms in the world, which poses many different demands. How to satisfy these demands is not specified, but rather it emerges through the way our interaction with the world shapes us. We need to integrate these mechanisms into our AI systems.

Do not count on “emergence” to save you

Also, I do not think that intelligence is something which will appear out of our system by some mystical emergence process–some kind of hope that predicts that if we amass enough computational complexity, then suddenly some kind of intelligence pops up and tries to achieve world dominance and wage war against mankind. These are things that need to be built explicitly into the systems, at least at some level. We have to identify the components which are responsible for things like personhood, social relationships, or what we define as being conscious. We have to take all these things and decompose them, and then, at some level we have to realize that as a system.

Taking the lessons: the MicroPsi architecture

Our own group has tried to take these lessons and come up with a relatively simple architecture, which we call MicroPsi. It is based on a theory by a German psychologist Dietrich Dörner. It uses a unified neuro-symbolic representation, integrates motivation into this system, and integrates autonomous agents which navigate virtual environments in order to pursue their goals.

The whole system consists of a neural network simulator for low-level perception, which is a special case in a more general semantic network for which you can define agents and run them. Then you have components which gives us multi-agent systems, so you can make many agents and have them interact. We have modules so we can implement the same structures on robots. The whole thing is built into a big architecture, and this theory asks what are the components of cognition, and what does it mean to be intelligent per se.

Motivation

This is based on unified modes of neuro-symbolic representations, which work by spreading activation and are hierarchical. The lowest level is connected to sensors and actuators. The whole thing is integrated with a motivational system, which learns by different kinds of pleasure and displeasure. It works by starting out with a basic set of predefined demands that the system has. Each of these demands corresponds to a drive–the drive to satisfy that demand.

The goals of the system are not predefined, but the demands of the system are given by the inherent structure of the system. These are of course physiological demands for nutrition and for physical integrity that have social demands. These shape our interactions with others and make us feel interested in others as agents, not as objects, and make us conform to social norms.

We also built cognitive demands into our system. For instance, the reduction of uncertainty, and to be competent in solving problems in general. These shape the exploration of the environment–not only the physical environment, but also our internal cognitive environment. Together they create a dynamic system which constantly changes its goals in order to build more complex representations of our environments to solve the tasks of the system. These tasks amount to the maintainance of the survival of the individual.

Neurosymbolic representations

f we look at the example of representation of that system, we have hierarchical representations. Just to give an example, you might be afraid of dogs because you may have been bitten once by one. On the other hand, you might like dogs because they are able to satisfy our demand for affiliation. This together creates meaningful representations of the environment which help us in a different context to only have those representations that are relevant to our achieving a goal.

MicroPsi is subject of ongoing work and is going to be discussed in more detail in the forthcoming book “Principles of Synthetic Intelligence” (Oxford University Press, 2008).

This entry was posted on Friday, May 16th, 2008 at 7:55 pm and is filed under AGIRI, Artificial Intelligence. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.