|
Inverse Icarus
|
 |
flying too low to the ground
May 2001
time: 00:28
|
|
quote:
Originally posted by Sava
I doubt it's possible to create a pure mechanical intelligence capable of the kind of thought processes that living beings are capable of. At most, I think that there would be cybernetic beings... biological brains and integrated cybernetical parts.
Computers, as we know them, aren't possible of intelligent reasoning and planning. And as advanced as they may get, they're just really powerful adding machines that do what a programmer tells them to do. |
WRONG. common misconception.
Biases and Heuristics: Gods Hard Coding, and Modular Intelligence
by Joseph Moskie
The study of the mind, and the related fields of human cognition, rational decision making, and artificial intelligence are relatively new fields in the realm of science. As these emerging fields have progressed, they have often mixed their methodology and their data, building on each other’s work. This interdependence has proven to be a double edged sword for these sciences, for although it allows the fields to advance relatively quickly compared to the older sciences, it often ends up with one of the fields waiting on the others for more information before it itself can advance. The field of artificial intelligence has come a long way in a relatively short period of time, but it can only go so far with our current understanding of the human mind, human cognition, and the rational decision making progress.
One of the most often used arguments against artificial intelligence is that essentially, no matter how far you break it down, you are “hard coding” rules into the artificial agent, which defeats the process of cognition itself. However, emerging research suggests that humans themselves base their own intelligence on simple rules that are innate in our very being. Kahneman and Tversky have proposed a system of “biases and heuristics” in human cognition that serve the same purpose that “hard coded” rules would serve in artificially intelligent agents, creating a basic cognitive framework from where more complex systems of intelligence and decision making could be developed. If basic human cognition is in fact based on a system of “biases and heuristics”, no matter how dynamic or malleable they may be, we can be justified in creating such systems in artificial agents, and building a more complex cognitive framework for the agent based on the bare-bone rules.
One should note the emphasis on the words “dynamic” and “malleable”, for the biases and heuristics in human thought probably won’t boil down to a simplistic if-then clause, ie. “If A then B”, but could end up being rather complex logical operations, with many interdependencies and conditionals. The claim that the biases and heuristics should be malleable implies that they should not be at all ridged; the set of rules governing cognition and decision-making should be easily controlled, easily influenced, able to adjust to changing circumstances. That is, to be completely frank, what cognition is all about.
Artificial Intelligence, as it has been presented to the general populous (in implementation, not research or theory) has come in three distinct forms: games, basic “intelligent” agents (PERI), and the apocalyptic death swarms portrayed in movies such as “Terminator”, “The Matrix”, and their kin. Ignoring the last one on the basis of “Hollywood v. Reality”, the other two have given off very skewed, very negative images of what artificial intelligence is and can be. The first, gaming, can be seen in blockbuster game hits such as “Sid Meier’s Civilization”, “Half Life”, etc, and also on a larger scale, ie. “Deep Blue”, the chess playing computer, that is sometimes able to beat its best human counterparts. In both cases, however, the artificial agents exist in a world with a very ridged set of rules, with very few variables (relative to a sentient being existing in the real world), and are focused on a very narrow set of tasks or goals. Deep Blue is an excellent chess player, yes, but Deep Blue did not struggle to learn how to play chess, in fact, Deep Blue is the brainchild of some of the best chess playing computer scientists in the world. Deep Blue has perpetuated the belief that artificial intelligence is all about “hard coding” the rules of a situation into the agent, and nothing more. When you boil it down, Deep Blue is not artificial intelligence, as it’s often passed off as; it’s simply a chess playing machine. It is not intelligent; it is skilled at its task. A seemingly “intelligent” task, but a task nonetheless.
Last semester I was exposed to an on-campus research project involving an “intelligent” robotic arm named PERI, an acronym for “Psychometric Experimental Robotic Intelligence”. This “intelligent” agent has been spawned from the brains of our own Dr. Selmer Bringsjord, the chair of Rensselaer’s Cognitive Science department, and a graduate student, Bettina Schimanski. At first glance, it seems remarkable; these people claim that PERI has successfully passed an IQ test, flawlessly. Then you read into it more, and find out that PERI has passed part of an IQ test, one section to be exact, and that the section it passed was rotating blocks to create patterns. Now, in my humble opinion, this seems more like a simple game than an act of pure intelligence. In fact, PERI even has the added bonus of being given all the information (ie, what images are on what side of the block) the nanosecond the test begins, whereas a human would have to actively explore the environment, and manipulate the blocks to discover what is where. Even after the human does that, PERI still has an advantage over the human, as PERI is allowed to label the sides and associate the image with the label, something the human is more than likely unable to keep track of in their head. It may sound speciest, to make such a claim (and I will address the issue of species centric thought later), but it comes down to the fact that PERI can win this “game” faster than a human can (Selmer brags that PERI can solve the puzzle in under a second, then takes quite a while to physically manipulate the blocks) doesn’t mean that PERI is capable of applying its skills to other cognitive operations. The two definitions of intelligence I have found that seem most concise are “having capacity for thought and reason, especially to a high degree”, and “exercising and showing good judgment”. In my personal connotation of intelligence, I put a lot of value on the conscious mind's flexibility, its versatility, the ability of the mind to apply itself to a multitude of tasks. In my opinion, any agent that can perform only one task successfully, no matter how “intelligent” a task it may be, cannot be said to possess intelligence.
We, as humans, have derived our denotations and connotations of the word “intelligence” solely on our species, and with good reason. For millennia, we have been the only “intelligent” being we are aware of, and human science is based on human experiences and human “fact”. This has been termed the “speciest fallacy”, or the “species centric fallacy”, whereas we, the “definitive example of intelligence”, want artificially intelligent agents to “jump through hoops”, and perform intelligent tasks on a level comparable to that of a human. One must look to the sensory devices many of today’s artificial agents have been given with which to perceive the world, somewhat crude when compared to the human sensory receptors. If we are to today say that artificial agents are not intelligent due to perceptual limitations that affect their cognitive processes negatively, then what will the artificial agents say when their sensory devices advance to a point much further than those of their human counterparts? As more advanced machines begin to evolve, we must look at them objectively when considering their intelligence, their sentience, despite the various shortcomings that may arise.
Now that I have described what Artificial Intelligence is today, how it is perceived, what it could one day evolve to, and how we should be objective when evaluating non-humanoid intelligence, it is time to describe exactly how an advanced artificial intelligence can be achieved. Daniel Kahneman and the late Amos Tversky had proposed a system of Biases and Heuristics, and applied their theory mainly to economics, especially in the areas of decision-making associated with risks. One of the classic examples they cite was preformed by Daniel Bernoulli (Bernoulli 1738), where he discovered that if you offered a person an 85% chance of winning $1,000 (with a 15% chance of winning nothing at all), or the option of definitely receiving $800, most of the people would choose the $800, despite the fact that .85 x $1,000 + .15 x $0 = $850, which is more than $800. This implies that there is something in human cognition that practices what he termed “risk aversion”, and he even went as far as to propose a function of “selective value”, or “utility”, is a concave function to money itself. Now, this is all well and good in the realm of economics, but again it appears that we are drawing nearer to a “specialized task”, which is the exact opposite of what Artificial Intelligence is supposed to be about, but that is only the appearance. Bernoulli’s concept of risk assessment, and what Kahneman and Tversky have added to it in the form of “Framing of Outcomes” can be abstracted to many non-economic problems relating to decision making in general, where several options are examined and thought of in terms of their possible outcomes. Kahneman and Tversky also go into many of the fallacies people commit, causing them to improperly weigh their options, which is a good read, but not directly related to artificial intelligence, unless your goal is to aid them in avoiding the fallacies innate in the human mind.
Regressing back to the original topic of Biases and Heuristics as proposed by Kahneman and Tversky, and taking it away from the economical context in which it was proposed, we end up with a theory that essentially states that we can take any abstract human cognitive process, and break it down continuously until we reach some system of rules that humans have “hard coded” into them. The problem with that theory is that it is extremely difficult to articulate exactly what that system of rules is, or how they work together to create the desired result. As stated above, the rules created by innate biases and heuristics are most likely extremely dynamic and malleable, completely adaptive to near any situation, and it is that fact that makes the rules so versatile, and at the same time so complex, and vicariously, so difficult to explore and define. What we have currently regarding cognitive frameworks is not much different than what we have had for many years now, a “Black Box” scenario where we know something is going on, but we simply cannot explain exactly what is happening. We can diagram it as Input à “Black Box” à Output. Although we currently cannot elaborate as to the exact functions of the box, we have proposed a theory that, after extensive study in the field, could eventually allow us to break down the “Black Box” even further into its basic components, its fundamental rules and functions. Once we have uncovered “the secret of the Black Box of human cognition”, artificial intelligence based on the same basic algorithms couldn’t be far behind.
Moving away from Kahneman and Tversky’s research on biases and heuristics, I’d like to discuss another issue regarding intelligence and thinking, namely modular intelligence. This has been discussed for many years under a variety of names (modular intelligence, levels of cognition, high/low-level cognition), and is still a hot topic in the fields of artificial intelligence and human cognition. The clearest way to describe what modular intelligence refers to is to first provide an example. The human sense of sight is a good example, and it has been professed by some to be modular intelligence, in the sense that the optic system itself makes decisions and acts intelligently independently of the brain and conscious thought. This can be perceived as “lower level thought”, where the optic system is “intelligent” to a lesser degree than the agent itself, and the agent unconsciously uses the “intelligence functions” of the optic system to reduce the amount of work that the conscious mind must deal with (J.J. Gibson, 1960). Gibson has argued that many visual properties (for example, texture) can be recognized by “low-level psychophysiological mechanisms, functioning without any control from high-level schemata or cerebral models”. This is truly a profound statement, with even more profound implications for artificial intelligence. This statement allows for a theory that essentially states that a human’s overall intelligence is actually the sum total of the work of smaller sub-systems that pass on filtered information to the conscious thought, which coincides with the fact that with artificial intelligence, there would be too much information coming in to deal with effectively. Rather than have the “main cognitive engine” of the artificially intelligent agent working on trying to “understand” the raw information from the environment, we can create “quasi-intelligent” sub-systems that are, as many professed “intelligent” systems are today, skilled at one task (or, optimally, good at a small number of specific, related, tasks). According to this theory, we would be justified in creating many interdependent sub-systems to manage and manipulate the raw data over and over again before the “main cognitive engine” even got a glance of it, forming the data into simplified, easily manipulated chunks, allowing the “main cognitive engine” to reserve the main CPU cycles for higher level cognition and decision making.
The emerging field of Artificial Intelligence is closely intertwined with the studies of the human mind, human cognition, and rational decision making, and will forever be. How can we claim to be able to go out and create artificial intelligence when we still cannot explain the basic foundations of our own cognitive frameworks? Simply put, we cannot. We must first strive to understand the mysteries of human cognition, to break down the “Black Box” into its basic components, whether they be biases, heuristics, or something completely different, and we must understand, before we can even begin to design a truly intelligent agent.
|
27-05-2003 05:36
- creation of God is the end of evolution. 
|