 |
AI Concept DocumentsConceptual Level Thinking |
|
|
|
| |
The Fourth Edition of the American Heritage Dictionary of the English Language in part defines intelligence in these ways: a.) The capacity to acquire and apply knowledge; and b.) The faculty of thought and reason. Few could argue with these handy definitions. But these are not technical enough to serve the need for formality in guiding AI researchers or those who should critique our work.
It’s not hard to acquire and apply knowledge, provided you open up the definition of "knowledge" enough. A web site can ask you for your name and probably quickly figure out where you live using it. That’s a pretty cool and practical application of a little knowledge. And while saying that intelligence is the faculty of thought and reason is also true, these terms are as opaque as the term they’re used to define. These kinds of shortcomings of such a definition are the main reason most people who consider the subject of AI answer the question of what "real intelligence" is are inclined to say they’ll know it when they see it.
I think that there is one word that sums up the meaning and threshold of the kind of intelligence AI strives for is "conceptualization", the ability to abstract useful meaning from the things it perceives in a flexible, generalized manner.
In the early days of AI, researchers decided to take on really heady problems like playing chess. They did a great job of making chess-playing machines, but these were not conceptualizing machines. They were simply built with the hard-coded gaming concepts of their makers. They went on to make critters that could navigate through rooms using complex maps and vision systems. These too were great at solving their own problem but were otherwise unable to do anything novel. Later beasts could do even better without maps or complex vision systems. These could even learn different behaviors to solve very real problems in locomotion and navigation, but they too were pretty dumb. The introduction of computers and robots that appear the empathize with human emotions are really no smarter, though they do help dispel the long-held notion that giving machines emotions will be the hardest problem to solve.
There has been a thread of AI research targeting the goal of a conceptual level thinking machine, if somewhat indirectly. The longstanding "Cyc" project of Doug Lenat is a prime example. Lenat’s work builds on a classic model of deductively reasoning expert systems -- if-then chains of logic. The key principle driving the Cyc project, now twice past its projected ten years, was to spoon feed Cyc huge volumes of information and let it ponder the relationships among it all until Cyc had enough knowledge to display the level of intelligence of a young child.
Some things trouble me about the Cyc approach, though. For one, Cyc doesn’t seem to be able to escape the world of the trivial by filtering out information that is of little importance to solving a problem. For humans, the more knowledge we have, the faster we are at solving problems. For Cyc, it seems, the more knowledge it has, the slower it gets at it. One can argue that it makes sense since the human brain is massively parallel, while the computer running Cyc is a serial processor, but I think that’s not essential. Humans don’t make important snap decisions on excessively lengthy chains of reasoning. They create "shortcuts" to speed the process. So can a computer program. Cyc’s problem here is more basic: it doesn’t really have a direction. It spends its time just pondering, largely at random. Even the most sedentary ivory-tower thinker doesn’t really do this. The missing key here is goal-directed thinking.
Another thing that seems to be missing is inductive reasoning. It’s innate for us to see patterns in things. One ambulance visits an apartment building and another one visits the same one an hour later and we presume there’s some common cause -- an infectious disease or serial killer -- behind it all. It may well simply be that one tenant had a heart attack and another one’s child drank some bleach a while later, but something so rare as multiple ambulance visits to one particular apartment building so close in time just screams "pattern" to us. Although one could argue that this is often a weakness and the essence of paranoia, it’s also clear that this sort of thinking is essential to our way of thinking. It’s seems likely that the same is true for most mammals as well.
In fairness to the Cyc project, it’s been a long time since I paid much attention to it, so some of the issues I’ve identified have been addressed, though I imagine if they had, the media would have taken far more notice than it has in following up with the project.
As evinced from icons of popular culture like Commander Data from Star Trek, many people consider the really hard challenge to be engendering emotions. I consider this the easy part. I also consider this an essential part of the human thought process. I don’t mean to suggest that emotions are a form of constructive thought of themselves. They are no substitute for good reasoning, but they surely modulate our thinking in as much as they help determine what we focus on at a given time. It’s hard to be inspired to think about the finer points of getting that second mortgage on your home when you’ve just witnessed a murder, for example. Emotions are one of those things that help to filter one’s focus so we don’t spend our days pondering the trivial relationships among mostly random concepts.
So the question remains: what is intelligence? Intelligence of the human sort is an assemblage of many things, but it is best recognized in one’s ability to derive long- and short-range courses of action based on the jumble of goals he has, which are ultimately derived from his values and the things he seeks to acquire in life, using a reasoning process and the ongoing integration of new information into an ever-growing hierarchy of concepts and evaluation of his actions by their fidelity to his goals and values.
That’s a mouth-full, but it is what I consider the essence. I’ll explain further below.
For those who want more background on the subject matter presented herein, I would point to some of the various non-fiction works by Objectivists -- especially Ayn Rand and Leonard Piekoff -- on the subjects of epistemology, capitalism, conceptualization, and so forth. While Rand and Piekoff have voiced skepticism of the study of AI, their ideas on these subjects are keenly analytical and relevant. For the fluid clarity of well-written words, I include some quotes here from Rand’s Introduction to Objectivist Epistemology, in which case I apply the label "ITOE" to identify the quotes. I also quote from her Philosophy: Who Needs It?, which I indicate with "PWNI".
Consciousness is demonstrated by an organism’s ability to react to its awareness of the world. An organism on a slope might move uphill toward sunlight, whereas a rock would more likely roll downhill from gravity. To add clarity to the subject, I’d say that an organism’s reaction differs from a simple physical reaction in that there is a level of abstraction involved -- some machinery that translates a literal sensation like the wiggling of an antenna into some kind of information that, after any measure of further processing, results in execution of an action, such as a leg kicking. The reaction does not have to be physically apparent, but nothing demonstrates consciousness better than a direct physical reaction to stimulus.
Some critics might say I’ve painted myself into a corner by defining consciousness in this way, since any computer can easily fit this criterion. So be it. I don’t exclude computers from consciousness. But I don’t think consciousness in this sense is the goal of AI. Intelligence presumes consciousness, but consciousness does not imply intelligence. There’s nothing special about being conscious.
Consider then the next level of intelligence: perception. Ayn Rand had this to say about perception in ITOE:
She distinguishes very purposefully between sensations and perceptions. Sensations deal with the literal impulses the senses fire off in response to recognizing whatever they are able to detect. Perceptions deal with a higher level of integration of those sensations.
Your eyes detect varying levels of light coming in from various directions. Your brain perceives a lake in the distant desert. The particular configuration of the points of light your eyes sense is a guaranteed given fact, while the perception of a lake is not a guaranteed fact. The "lake" may be a mirage, but the fact is that the light that hits your eyes in such a way that makes you perceive a lake is to your eyes identical to the light that would come from a real lake. Still, you are entirely unaware of the light levels; you are aware only of the image suggesting a lake in the distant desert itself.
Some of the most basic AL experiments present organisms at a sensory level of intelligence -- what I call "knee-jerk behavior". Conway’s life is a great example. An organism, existing as a square on a grid, decides what color it will be in the next step based on what combination of colors its nearest neighbors are in this step. My wife and I made a robot as a project for our circuits lab in school we called "Kikucci" which also illustrated the idea. With two photodetector eyes, two motorized wheels, and some simple electronics, we were able to make Kikucci drive towards a bright light source -- an entirely sense-and-response behavior set.
One could ascribe perception to both these experiments. In the Kikucci experiment, we could have defined a flashlight as food and the behavior of moving toward the light as seeking food. We would then conclude that Kikucci could perceive food through its limited senses. This is a bit contrived, since the robot obviously couldn’t eat the flashlight (wouldn’t that be cool?) to use it for sustenance and so it wasn’t really food. Still, this helps illustrate the meaning of perception. It also leads us to one very important premise I hold in regard to the subject of AI: that intelligence is what intelligence does. It is nearly pointless to argue over whether Kikucci was intelligent. It was obviously not very bright, but it clearly had an awareness of the world and responded to its awareness. Regardless of the significance you ascribe to the behavior of driving toward a bright light, it’s obvious the behavior was there and was driven by an abstract translation of sensations into action, regardless of what you think it perceived.
Almost all AI research proceeds in this way. Perception is rarely spoken of but always assumed. A program that responds with a cute repertoire to human facial expressions will generally be built with at least two layers: one to pick a face out of a video feed and interpret its facial expression and another to respond to the perceived expression. A robot that drives a car has one layer to interpret pixels in an image as lines on the road and another layer to respond to the perception that the road is curving to the left ahead.
I see emotions as being entirely mechanistic and understandable. To quote from Ayn Rand’s PWNI:
I distinguish carefully between sensations -- as in physical pain or pleasure -- and emotions. The fundamental difference is that sensations are direct and emotions are driven indirectly by perception. Your hand touches a burner and the sensation of pain is driven without further ado by sensors in your hand that directly perceive heat-pain. In many cases, strong pain or pleasure sensations automatically lead to reactionary behavior -- moving your hand away from the hot burner, for instance -- though this is not guaranteed.
Emotions are very secondary, by contrast. To explain why, let me start by identifying the primary emotions all mammals and most other "higher" animals experience:
Yes, you can quickly find hundreds of words we use in everyday parlance to identify distinct "emotions", but I argue that most of them refer not to emotions, per se, but some combination of these primary emotions with their causes or responses. For example, envy can be defined as anger because someone else has something one desires. Embarrassment could be defined in terms of the fear and anger one feels because his peers become aware of something that lowers his standing in their eyes.
One essential component of emotions is that they are automatic responses that are inborn and unchangeable at the lowest level. Emotions are always responses to perceptions. Joy is always driven by the recognition of the acquisition -- or certainty of future acquisition -- of something of great value. Fear is always driven by the recognition of a potential threat to what one values. Sadness is always driven by the recognition that some value has been lost or some opportunity missed. Anger is driven by the same; the subtle difference is that whereas sadness generally leads to resignation to the loss, anger leans in the direction of action to regain the value or to avoid further loss.
As such, emotions are responses to their precipitating perceptions. The most obvious responses for biotic organisms like mammals are releases of related hormones. Some hormones are intended to help strengthen the brain’s ability to remember the significance of events that cause strong emotions. Anger releases hormones that tend to make the mind focus immediately on resolving the perceived problem and in some cases to act immediately in ways that the non-angered mind might not act for fear of associated dangers. Emotions as such act as general modulators of how our brains function -- in essentially the same way that psychoactive drugs do.
If it’s not clear already, let me drive a point home. Emotions are not a form of intelligence or knowledge. Emotions are always responses to what one perceives. They may drive one to think in a certain way, but they are not thoughts as such. Nor does feeling good about a thing make it good, nor vice-versa.
Which leads me to the next essential point: while the immediate, low-level cause of an emotion is hard-wired and unchanging, the higher-level abstract cause of an emotion is not. If joy is in response to perceiving the receipt of something of value, then it must be pointed out that what one perceives as value is variable. For example, you are born perceiving the receipt of something that tastes, smells, or feels good as you acquiring something of value. Naturally, you feel joy when you receive something that feels particularly good, then. With a certain change in your way of viewing the world, you can quickly train yourself to perceive receiving, say, some good-tasting object -- such as sugar-laced rat poison -- as something that is bad for you. When you taste sugar you think may contain rat poison thereafter, you will more likely feel fear than joy. That’s an emotional response "reprogrammed".
In the context of AI, consider an artificial brain endowed with emotions in this way. It has one whole layer devoted to perceiving objects and states of being. It has another layer that has emotional responses that are wired to a handful of those perceptions devoted to states like receipt or loss of value. There are then patch cables connecting other perceptions such as the taste of sugar with the appropriate perceptions associated with those value-judgment perceptions. Those patch cables may be rearranged through a learning process.
Technically, the learning part is not essential to isolating emotions from other mental components, but it becomes far more difficult to separate emotions from physical sensations like pain when you discount the "reprogrammability" facet of emotions.
All living organisms require certain material objects to survive or reproduce, which are two defining characteristics of life. (Let me point out here that an intelligent organism doesn’t have to be alive, since reproduction is essentially irrelevant to intelligence.) Those material goods, such as water or sunlight, are said to have value to such an organism. Some organisms seek certain states of being such as joy. Those actions that lead to such states also have value. Action to avoid objects, actions, and such that cause one to lose or fail to gain value is also itself a value. In fact, all objects, actions, and states of being can have attributions of value associated with them in the context of any given organism. That valuation isn’t necessarily the same for all organisms. A chocolate bar, which can kill a cat, is obviously not of value to cats, while it is of value to many humans.
Value is measurable. The value of one thing is measured in relation to the value of another as greater, lesser, or equal to. As such, value measurement can get extraordinarily complicated.
Value is perceived. Most organisms are pre-wired to value anything that promotes good basic health and reproduction. Still, mammals can easily be moved to value certain psychoactive drugs despite their danger to health and reproduction. Assignment of value is like a patch board with wires connecting value assessments to each mental abstraction. And tying it back to emotions, the acquisition or loss of something of value can lead automatically to an appropriate emotional response.
Every action taken by an organism can be said to be driven by some goal. The goal is always to acquire or avoid losing some particular value or set of values. Since non-organisms aren’t conscious, they can’t value things, let alone set goals. Thus, goal-directedness is a defining attribute of organic action.
The ability of an organism to change its relative evaluations of objects, actions, and states of being through a ongoing and conscious evaluation process is what defines volition, or "free will". Make no mistake on this point, though: there is no contradiction between the notion of mechanistic predictability and free will. Free will is not defined by any divine or otherwise unpredictability. If it were, a coin could be said to have free will about whether it will fall as heads or tails on a coin toss. Volition is entirely about an organism being able to make one value judgment in one circumstance that may be completely different from another judgment in the exact same circumstance later because its context of knowledge has changed in the interim.
 |
Produced in cooperation with Carnell Information Systems, Inc. |