"We make the relevant things in the current situation salient, by whatever makes for salience."
- Daniel Dennet

Introduction
A defining characteristic of intelligent agents is their ability to choose between multiple live possibilities. In fact, in choosing one option to the exclusion of others, intelligent agents generate information. In writing this essay I am choosing certain words from the English language and in doing so generate information by eliminating possibilities. I am conveying information to the reader by reducing uncertainty. At each position in this sentence I am selecting from the reference class of my English vocabulary and choosing just one word. Those words that I am aware of are part of my general informational context: my complete representation of the world.

An intelligent agent works from within a certain representation of the world. Much of the effort in Artificial Intelligence has been put into developing a proper representational system which the AI agent will use to model the world within which it is situated. One of the big problems in AI known as the Frame Problem is related to this need to find an efficient, robust, stable, simple and general representational system. Specifically, the Frame Problem addresses our need to deal with what Murray Shannahan calls "the common sense law of inertia": the fact that most things tend to stay the same given any particular event. If I paint my car, I do not need to consider whether this will cause the toilet in my house to disappear. In essence, the Frame Problem is concerned with defining laws of non-change within a representational system of the world.

Daniel Dennet sums up the problem in his wonderful little essay "Cognitive Wheels: The Frame Problem of AI" with a story of three robots: R1, R1D1, and R2D1. In all three examples the robots are looking in a closet for their battery source. R1 goes to the closet, sees that his battery source is on a wagon, also sees that a ticking bomb is on the wagon, but can not deduce that by pulling out the wagon with his battery source, he will also be pulling out the bomb. So, while R1 has been programmed to pull its battery source out of the closet, it has only the capacity to recognize that a bomb is on the wagon and not the implications of this fact, and the bomb subsequently blows R1 into bits. R1D1 on the other hand is able to consider the implications and side-effects of its acts. However, he gets easily bogged down in an unending web of deductions, and just as he is considering that pulling the wagon will not change the color of walls in the room, he joins R1 in robot heaven. R2D1, being the most advanced robot of all, is able to distinguish between relevant and irrelevant implications by tagging them as such. Unfortunately, R2D1 spends far too much time ignoring thousands of irrelevant implications as the bomb explodes and his designers go back to the drawing board.

The reason I highlight Dennet's description of the Frame Problem is that it brings forth a fundamental requirement for temporally relevant intelligent agents. In order for R2D1 to have a temporally relevant intelligence, it would need to be able to focus on the relevant information in real time. Human beings possess the ability to frame their information in such a way that the relevant information becomes salient in their conciousness: automatically. There is no flagging of irrelevant information. Rather, there is some fundamental mechanism by which human beings hone in on a specific set of information that is relevant at any particular moment.

It is from within this representational system that an intelligent agent makes decisions based on its wants, beliefs, desires, knowledge etc. When a person makes a decision, she does so from within her representational system. Human beings have the same basic underlying framework/hardware for representing the world. However, the distinguishing feature between representations is the contextual knowledge that each person uses which depends on our particular histories, experiences, assumptions, values, current circumstances and interests. We can call all the knowledge that an intelligent agent has about the world its general informational context. With each cognitive act, the agent must draw from this general informational context to plan, to make decisions, and to act: to be an intentional agent.

One of the great problems in Artificial Intelligence involves the question of how an agent moves from its general informational context to a more specific, circumstantial context. How does an intelligent agent narrow down the relevant knowledge from the irrelevant knowledge. At any point in time, whether in a discussion or in grabbing a beer from the fridge, a person ignores the majority of information in its general informational context, only taking into account the specific knowledge relevant to a particular circumstance. When I'm having a discussion about terrorism, my knowledge about ice cream is probably irrelevant and not taken into account.

I. General Informational Context

The general informational context is the complete set of information that an intelligent agent has about the world at any particular time. Developing AI systems with extremely large general informational contexts has not been difficult. We currently have the technology for massive information storage. The problem, of course, has been how to properly use this information within situated contexts. If a robot has access to all the knowable information about cars, beer, and terrorism, how can the robot know which information is relevant in a particular situation? Developing expert systems has proved to be one of the most fruitful AI research projects. Expert systems however come short on two ends when it comes to our understanding of general intelligent systems. First of all, the word "expert" is very telling: these systems have a very limited and specified range of knowledge. Second, they merely supplement and complement human knowledge. They are dependent systems which require specific instructions for diagnostics and which require human beings for interpretation, discernment, and implementation. The key here, though, is that although a particular expert system may be able to diagnose an illness, it would not know the first thing about how to drive a car or who the President of the United States is. The expert system is, to put it bluntly, stuck in a mental rut.

Human beings, on the other hand have huge general informational contexts which include a large variety of information about the world. The problem in AI in relation to human-like general informational contexts has been how to categorize this information into what I will call specific informational contexts. What information is relevant to a particular conversation or task? When I'm putting together a puzzle, I need not focus on the architecture of the house, and I certainly don't need to consider whether my fitting together a piece of the puzzle will cause my house to fall down. Human beings have the ability to consider only the relevant information in a particular situation. They are able to seamlessly and dynamically identify specific informational contexts.

II. Salience and Specific Informational Contexts

The specific informational context (SIC) is simply a subset of an intelligent agent's entire body of knowledge: the general informational context. It is all of the background knowledge and experience that an agent has in reference to a specific situation. If we are talking about cars, then the SIC is all of the information a person has stored in her mind that is relevant to cars: parts of cars, types of cars, how car engines work, etc. Interestingly, in an expert system there is only one SIC and it is identical to the general information context. If we have an expert system that deals with cars, then we could say that its specific informational context is pre-identified as its general informational context. As noted above, this is a case of psychological inertia in which an intelligent agent can not dynamically extend or combine specific informational contexts within a larger general informational context. In humans, on the other hand, each SIC is situated within the larger general informational context. A big question for those of us in Philosophy of Mind and Artificial Intelligence involves how an intelligent agent is able to move between specific informational contexts within the general informational context. How is the mind able to seamlessly and effortlessly situate its knowledge within a specific context? Could we find a mechanism to perform the same function in AI systems?

The specific informational context finds its analogue with the reference class in statistics. In statistics, the reference class is the set of all relevant possibilities. The reference class is ultimately independent of physical reality: it is a conceptual tool that the statistician uses to analyze possibilities. The key point in this analogy is that just as in statistics, where the statistician must identify the reference class, it seems that the intelligent agent must also identify the specific informational context in which it operates at any given time. In human beings this identification happens without much (if any) effort. When conscious, a human being is always working within a specific informational context.

III. Specifications
Specification involves the ability to identify a small set of information which meets a functional need at any given time. Specification occurs when we identify a set of words that would be appropriate to use at any given position in a sentence. It is also the method that intelligent agents use for problem solving. Often times a specification happens without much deliberation. Other times it requires very focused concentration for long periods of time.

IV. Choice/Event

V. Action

VI. Identifying the Specific Informational Context and Relevant/Salient Information

VII. Updating the General Informational Context

Learning and updating occur at the level of general informational context. The reason for this should be evident from the nature of specific informational contexts: they are not rigid sets with discreet borders. Specific informational contexts are dynamically identified and made salient in our consciousness, but it is not as if they are pre-identified. Borrowing from computer programming terminology, we could say that while information stored on your computer in files is pre-determined, specific informational contexts are identified in real-time at run-time. Programming information into a robot as a large set of "scripts", "frames", or "stereotypes" (ie. categorizations) will not solve the problem. Rather it would reduce a dynamic mechanism to static block of information. Specific informational contexts are identified on an ongoing basis, by dynamically retrieving and organizing relevant information in real-time. The task of the AI theorist is to develop a representational system that is flexible and dynamic on which identification mechanisms can operate.

VIII. Conclusion

Intelligent agents require the ability to store information about the world in a general informational context. The general informational context is the set of all knowledge, beliefs, and desires that an agent has. However, having just this general informational context is not enough. It appears that human cognition requires the ability to seamlessly identify and operate on a more specified informational context that is relevant to a particular context. Furthermore, after situating itself in a specific informational context, the intelligent agent must identify a set of relevant and salient information from which to choose (and thus generate information). After identifying this target the intelligent agent must then finally make one last choice: which piece of information to act on.

Your post can't help but remind me of gestalt psychology. First, there is the gestalt of Perls. He views the human organism as a gestalt, a unitary structure, that continually needs to maintain its unity.

An individual caught in the desert and facing dehydration will immediate focus attention on the source of water in the distance and overlook the pot of gold nearby. If it is past noon and I haven't had my morning cup of coffee, I will hurry to the local coffee shop. Since I'm not looking for a job, I will not notice the help wanted sign in the window. When I get inside I will also overlook the half-price sale on ice cream. Ice cream contains no caffeine, but I NEED my cup of coffee. So we have a remarkable ability to filter out unnecessary information and focus on what we need at the moment. Though a short time later, we focus on something else as our needs and desires have changed; and the things needed to "complete the gestalt," whether a biological or psychic gestalt, have now changed.

So our actions and our attentions are a matter of our unitary nature. And wholeness is also an important feature of the more purely gestalt school of Wertheimer and Kohler as well. We discern unitary structures. For human perception, a straight line that forms the side of a rectangle is not the same thing as a straight line in isolation. And of course, we're all familiar with the gestalt "pictures" that can reveal one thing and then another depending on the perspective with which we view them, or the Necker cube that can be seen as an empty box or as a solid block.

So it seems that this "wholeness" of human perception can act as both a center of focus and as a filter simultaneously, and as the example from Perls suggests, this "perception" doesn't need to be limited to vision but includes the whole body including the other senses and its internal function, both physical and psychic. So it seems to me that AI must find some way to build in this holistic factor, or rather, a range of overlapping holistic factors. (Could holography play a part here? Could it be that current hardware isn't up to the task? I know very little about holography or computers so I'm just throwing that out).

However, let me add my own bias. I don't think it's possible. The whole problem presupposes materialism. Some time ago I was cogitating on the "zombie problem" presented by David Chalmers in his book, "The Conscious Mind." It led me to thinking about the function of consciousness and I wrote a little essay which I didn't do anything with. In the essay I imagined a little baby zombie and his real-life "brother." They were capable of doing the same thing and both had the same knowledge. So they could both suckle at their mother's breast and both knew that her body was warm, both could explore their environments, etc. But while both could feed, seek warmth, and explore their environments; only one would want to. Of course, the zombie would die without feeding, but why would he care? It is sense perceptions that give the real baby a reason for living. It is sense perceptions that drive our desire to survive. And these same sense perceptions underly our ability to form gestalts. We need meaningful associations. Gestalts provide a context for meaning but he meaning is ultimately important to us because of sense perceptions. Sense perceptions are what we are, and why we want to go on living.

So I think this completely refutes materialism. Not in a logical sense of course, but in a practical and meaningful sense. In fact, materialism has no place for meaning. But I think this may be a reason why AI will never solve the framing problem. Once r2d1 has figured out that there's a bomb in his wagon, why should he do anything about it?

quote:

---

Originally posted by Micah Sparacio:
[QB

At any point in time, whether in a discussion or in grabbing a beer from the fridge, a person ignores the majority of information in its general informational context, only taking into account the specific knowledge relevant to a particular circumstance. When I'm having a discussion about terrorism, my knowledge about ice cream is probably irrelevant and not taken into account.

QB]

---

I think even this portrayal of human intellect is lacking. There seems to be more involved than just narrowing the choices from the general to the specific context. A few examples come to mind.

First, we don't seem always to be very proficient at considering all relevant data within our general field of information. We call the results "unintended consequences, or oversights." How often do you find yourself saying "I should have known better!" Perhaps there is a relationship between the apparent "sloppiness" of our thinking, and our ability to think abstractly. Second, when engaged in creative work, it is the ability to "think outside the box," or to consider connections from a given problem context, to seemingly unrelated contextual fields. Call it the "McGuiver" approach if you will. Third, there are the accidental discoveries, such as in the cases of Teflon and Post-It Notes. Finally, emotional stress as a motivator cannot be overlooked. Almond Stroger invented the "Stroger Switch" thereby making dial telephones possible, because the Bell operator was diverting calls to his undertaking business to her brother-in-law.

[ 26 February 2002: Message edited by: marcus ]

I agree with your assessment that needs, desires and beliefs are the things that make general human intelligence so hard to model. In fact, these things are directly related to our identification of saliency: what things stick out in relation to our current needs, wants, and beliefs.

I also think that the points you bring out are directly related to my sense that the ideas I presented are incomplete. Human intelligence is, after all, more than syntactic information processing and retrieval. However, my intention is to introduce a model by which we can at least explore the potential for an AI mechanism of saliency and specification. I think the wrong way to go about it is to define "systems of ignorance", which is my impression of where frame problem solutions have gone in the past (this is definitely not how humans go about it). More on target is the idea that there may be pre-defined but flexible categories of information that we situate ourselves in (frames, stereotypes, etc). However I still think that this is non-human. Of course, AI does not have to commit itself to human like intelligence...but after all, that has been the vision over the horizon for the past 50 or so years.

Ultimately I am interested in the information flow that exists in the human process of specification-decision-action. The flow of functional information from the human being into the physical world. I'm interested in the causal specificity (the mechanisms by which it happens), the different ways in which the information becomes manifested in the world, and the routes of information flow. Even if this capacity is not possible in AI systems, the mental exercise is ideal because it causes us to focus on certain features of our cognition that we may ordinarily look past when concerned with the wholeness of consciousness

Also, one of the ideas that first came to me was that the specific informational contexts are NOT static but are dynamic. This is what allows for "thinking outside the box". In fact, AI researches have proposed in the past that humans think in "stereotypes." This is often correct, we think in patterns, and often are unable to break outside of the patterns of thought we've been conditioned to. However, there are rare occassions when people are "thinking outside the box" and I hope to emphasize this in subsequent writing. Static categories of thought won't do. For creative thinking, you need to be able to mix and match, stretch and skew the specific informational context.

quote:

---

Ultimately I am interested in the information flow that exists in the human process of specification-decision-action. The flow of functional information from the human being into the physical world. I'm interested in the causal specificity (the mechanisms by which it happens), the different ways in which the information becomes manifested in the world, and the routes of information flow.

---

It seems to me from what little I've read on this subject that the focus has been too tightly wrapped up in the idea of explaining human consciousness in terms of information processing and too little attention has been paid to the type of information that is being processed. A computer processes digital information. So does the human brain. But there is something remarkable about the human consciousness. All of a sudden, the information that arrived in digital form suddenly becomes transformed into an analog form. So I know where the door to my apartment is because I have the information in my brain. A computer, of course, could also know that. But I don't just know it; I see it. And I don't see it in my head. I see it where it is in the space about me. So my mind is able to represent the world about me in space. And it is representing the space itself as well. So the mind isn't even in space. It represents the space itself.

And this is where the materialist model breaks down. How does this happen? The only materialist model I can think of that might be like this is holography. So maybe materialists can come up with an answer there. I don't know.

At any rate, it seems to me that the mind is a kind of logical space that tells me about the world around me but does so in a very meaningful from. In fact, it is so meaningful that I can't really consider myself to be the nerves and brain and c-fibers that intercept this information. I am, first and foremost, the information itself.

But now there seems to be an anomoly here. This information that is me is much richer than the mere digital information that I would receive if I were a computer. But information is supposed to be something that limits choices and possibilities. So memory must be very important here. I touch a hot stove and it burns. I quickly remove my hand. But this can also be achieved as easily by an automatism. Still, I learn to be careful around a hot stove. A robot could be programmed to learn the same thing, but it wouldn't be nearly as efficient in programming space. So our sentient experience we appear to be more efficient, but I think there's more to it than that.

So this analog information, sentient experience, limits my options by building up memories. But it doesn't foreclose options as completely as a computer program would. So I seem to end up with the restrictions that information requires but still possess the freedom to override those restrictions. I can play football or be a boxer in spite of the pain that those things involve. So that seems to be heading toward something like free will. But I really haven't carried it any furhter than that. I don't know if this is helpful or not. It's just some of my musings.

quote:

---

BTW, Marcus, you mentioned a paper you wrote on the Zombie problem. I'd be very interested in reading it either via email or via a post to this discussion board.

---

quote:

---

Originally posted by marcus:
Marcus writes:

At any rate, it seems to me that the mind is a kind of logical space that tells me about the world around me but does so in a very meaningful from. In fact, it is so meaningful that I can't really consider myself to be the nerves and brain and c-fibers that intercept this information. I am, first and foremost, the information itself.

.

---

quote:

---

Marcus: Do you really think that the agent we refer to when we say "I" is "first and formost, the information itself?" If so, then why do we always construct our sentences as if we were judges, arbiters, evaluators of that information.

---