Artificial Intelligence and Self-organizing Structures
Thesis: Unlike the AI systems and computer technologies we rely on today, the most powerful AI systems of the future will adapt, learn, self-organize and emerge as intelligent, much like the intelligent systems we observe in nature.
In a recent post, I talked about the differences between wisdom and intelligence as they apply to the human intellect. Next I would like to discuss the implications for technology, more specifically artificial intelligence. Many leading authorities believe that true artificial intelligence, as seen in the quality of the human intellect, will require impeccably simple, beautiful algorithms capable of advanced pattern-recognition and learning, rather than the fully-programmed, detailed designs often used for modern applications. These algorithms will result in emergent intelligence.
Pattern Recognition in A.I.: Seeing Shakespeare?
Although we have created computer systems that use pattern recognition, they are built for very specific purposes. For instance, I'm sure your familiar with the SAT. The College Board recently added a writing section to the SAT, similar to sections that appear on their graduate school tests . Believe it or not, test taker's essays are graded by a computer. Both a human and a computer grade the essay, and in the rare case that the computer-generated grade differs from the human grade, a second human grader acts as the tie-breaker. The system works on pattern recognition. The computer compares a given essay to an extensive database of good, mediocre, and bad essays for vocabulary, grammar, style, and content. By identifying the previously graded essays that most-closely resemble the candidate essay, the computer almost always yields a score within one increment of a human grader. Really, the computer is doing just what the human is doing when they blow through 100 essays a day, using the gut instinct and intuition developed from grading thousands of essays to quickly assign a without an in-depth analysis. This program, while impressive, is limited to a handful of applications. We have developed programs for recognizing faces, text, sounds, and other things, but never a single, unified system that relies on the same architecture and programming to recognize and interact with the plethora of objects humans encounter.
Check Mate?
In tasks less fuzzy than grading essays, with clearly defined rules and goals, brute processing has been used to overcome the cognitive abilities of humans. A few years ago, IBM developed the first computer (Deep Blue) to defeat the reigning international chess champion. Deep Blue calculates an exhaustive number of potential scenarios, looking as many moves as possible into the future, and using sophisticated rules to prune away the scenarios that are less relevant. For each and every one of these scenarios, Deep Blue is thinking through each move. When Gary Kasperov plays chess (the champion) he does not consciously think through billions of individual moves. He relies mostly on experience and intuition to recognize patterns that guide him to his next move. Kasperov's style might be described as more "creative" than that of Deep Blue. And in reality, human situations require more creativity than the black-and-white rule book of chess. Raw computing power used for the calculation of scenarios won't write an endearing, Shakespearan love note or reproduce the human penchant for good irony and laughter. While some of the A.I. systems to-date are impressive, on an individual basis, they cannot begin to compare to humans in their adaptability and versatility.
Decentralized, Self-organizing Structures
The physical differences between humans and today's computers explains a lot. The brain does not do the type of thinking Deep Blue does. Computers rely on the precise calculations of a single processor, a Ferarri of computing; whereas the brain relies on billions of less-powerful processors working in parallel to arrive at a sometimes imprecise but often more-brilliant result. This works by synapses, or the processors, in your brain constantly firing and sensing others around it firing in a complicated network. This is decentralized processing. The value of the brain's decentralized processing power can be seen in other less mysterious examples, like the efficiency of markets and the intelligent behavior of ant colonies.
In much the same way your brain synapses work in concert, there are hundreds of millions of people participating in the public markets, each registering their knowledge and processing power through the buying and selling of securities. For everybody buying a security, there's some other guy who thinks it's time to sell, yet despite this messy system of seemingly incompetent and imperfect individuals, we have a fairly efficient market driven by Adam Smith's "magic hand." Similarly, ant colonies build extremely complex structures, including bridges, despite the individual ants knowing nothing about the global system and each following only a few simple rules on a local basis. Systems such as these, including the human brain, are often referred to as emergent intelligences. Just think about relative failure of centrally planned communist economies to capitalism-based economies and the concept of emergent intelligence opens up a whole new realm of philosophical debate.
Now, lets discuss how its possible to achieve self-organized intelligence. This type of intelligence is by definition emergent, and I would argue can only occur as such (just as no one human has the intellectual capacity to devise the computer program I discussed above). Think about the delicate and complicated nature of the human brain. Right now you are a sensible person who functions compatibly with society (I hope) as a result of the billions of molecules in your brain being positioned and interacting exactly as they are currently. Now, if I expose you to a couple weeks of Chinese water torture (forcing you to endure constant, evenly-tempoed dripping on your face), you will in theory go insane. How much did the physical structure of your brain change?--not much, probably less than 1%, and yet you are no longer an "intelligent" being capable of functioning in society. Similarly, over time markets have self-organized from a primitive barter economy into an extremely complex and efficient capitalist structure that is reflective of an intelligent civilization. This structure has hundreds of thousands of different companies and billions of entities, all with different roles, each relying on thousands of processes and channels of communication to operate together in an intelligent manner. No one man is entirely essential to the whole, and can be replaced by another similarly trained individual, much like no one cell is essential to the brain (in fact every 7 years on average all the cells in your body is replaced through natural processes; its the structure that matters). But if you were to change just a few things, like all the pass codes for wire transfers, records of security ownership, and redistribute anybody who knows anything about running the Federal Reserve to basket weaving in Peru, the system would literally go haywire.
The point is, there's practically an infinite number of wrong combinations of how to organize billions of brain molecules (or billions of people in a civilization) for every right way that achieves intelligence and compatibility with an environment, and consequently, a truly intelligent computer system will have to learn and develop naturally from its environment in some capacity, and then perhaps be copied. Thus, human-like artificial intelligence will require design that centers on pattern recognition and learning. This all may sound like bad news for AI, but I think it bodes well. It means that an extremely complex, intelligent system can develop from the infancy of a few brilliant algorithms designed to achieve perpetual, adaptive self-organization through learning. We just need a few brilliant minds to discern the nascent self-organizing automata that underlie "intelligence."
In a recent post, I talked about the differences between wisdom and intelligence as they apply to the human intellect. Next I would like to discuss the implications for technology, more specifically artificial intelligence. Many leading authorities believe that true artificial intelligence, as seen in the quality of the human intellect, will require impeccably simple, beautiful algorithms capable of advanced pattern-recognition and learning, rather than the fully-programmed, detailed designs often used for modern applications. These algorithms will result in emergent intelligence.
Pattern Recognition in A.I.: Seeing Shakespeare?
Although we have created computer systems that use pattern recognition, they are built for very specific purposes. For instance, I'm sure your familiar with the SAT. The College Board recently added a writing section to the SAT, similar to sections that appear on their graduate school tests . Believe it or not, test taker's essays are graded by a computer. Both a human and a computer grade the essay, and in the rare case that the computer-generated grade differs from the human grade, a second human grader acts as the tie-breaker. The system works on pattern recognition. The computer compares a given essay to an extensive database of good, mediocre, and bad essays for vocabulary, grammar, style, and content. By identifying the previously graded essays that most-closely resemble the candidate essay, the computer almost always yields a score within one increment of a human grader. Really, the computer is doing just what the human is doing when they blow through 100 essays a day, using the gut instinct and intuition developed from grading thousands of essays to quickly assign a without an in-depth analysis. This program, while impressive, is limited to a handful of applications. We have developed programs for recognizing faces, text, sounds, and other things, but never a single, unified system that relies on the same architecture and programming to recognize and interact with the plethora of objects humans encounter.
Check Mate?
In tasks less fuzzy than grading essays, with clearly defined rules and goals, brute processing has been used to overcome the cognitive abilities of humans. A few years ago, IBM developed the first computer (Deep Blue) to defeat the reigning international chess champion. Deep Blue calculates an exhaustive number of potential scenarios, looking as many moves as possible into the future, and using sophisticated rules to prune away the scenarios that are less relevant. For each and every one of these scenarios, Deep Blue is thinking through each move. When Gary Kasperov plays chess (the champion) he does not consciously think through billions of individual moves. He relies mostly on experience and intuition to recognize patterns that guide him to his next move. Kasperov's style might be described as more "creative" than that of Deep Blue. And in reality, human situations require more creativity than the black-and-white rule book of chess. Raw computing power used for the calculation of scenarios won't write an endearing, Shakespearan love note or reproduce the human penchant for good irony and laughter. While some of the A.I. systems to-date are impressive, on an individual basis, they cannot begin to compare to humans in their adaptability and versatility.
Decentralized, Self-organizing Structures
The physical differences between humans and today's computers explains a lot. The brain does not do the type of thinking Deep Blue does. Computers rely on the precise calculations of a single processor, a Ferarri of computing; whereas the brain relies on billions of less-powerful processors working in parallel to arrive at a sometimes imprecise but often more-brilliant result. This works by synapses, or the processors, in your brain constantly firing and sensing others around it firing in a complicated network. This is decentralized processing. The value of the brain's decentralized processing power can be seen in other less mysterious examples, like the efficiency of markets and the intelligent behavior of ant colonies.
In much the same way your brain synapses work in concert, there are hundreds of millions of people participating in the public markets, each registering their knowledge and processing power through the buying and selling of securities. For everybody buying a security, there's some other guy who thinks it's time to sell, yet despite this messy system of seemingly incompetent and imperfect individuals, we have a fairly efficient market driven by Adam Smith's "magic hand." Similarly, ant colonies build extremely complex structures, including bridges, despite the individual ants knowing nothing about the global system and each following only a few simple rules on a local basis. Systems such as these, including the human brain, are often referred to as emergent intelligences. Just think about relative failure of centrally planned communist economies to capitalism-based economies and the concept of emergent intelligence opens up a whole new realm of philosophical debate.
Now, lets discuss how its possible to achieve self-organized intelligence. This type of intelligence is by definition emergent, and I would argue can only occur as such (just as no one human has the intellectual capacity to devise the computer program I discussed above). Think about the delicate and complicated nature of the human brain. Right now you are a sensible person who functions compatibly with society (I hope) as a result of the billions of molecules in your brain being positioned and interacting exactly as they are currently. Now, if I expose you to a couple weeks of Chinese water torture (forcing you to endure constant, evenly-tempoed dripping on your face), you will in theory go insane. How much did the physical structure of your brain change?--not much, probably less than 1%, and yet you are no longer an "intelligent" being capable of functioning in society. Similarly, over time markets have self-organized from a primitive barter economy into an extremely complex and efficient capitalist structure that is reflective of an intelligent civilization. This structure has hundreds of thousands of different companies and billions of entities, all with different roles, each relying on thousands of processes and channels of communication to operate together in an intelligent manner. No one man is entirely essential to the whole, and can be replaced by another similarly trained individual, much like no one cell is essential to the brain (in fact every 7 years on average all the cells in your body is replaced through natural processes; its the structure that matters). But if you were to change just a few things, like all the pass codes for wire transfers, records of security ownership, and redistribute anybody who knows anything about running the Federal Reserve to basket weaving in Peru, the system would literally go haywire.
The point is, there's practically an infinite number of wrong combinations of how to organize billions of brain molecules (or billions of people in a civilization) for every right way that achieves intelligence and compatibility with an environment, and consequently, a truly intelligent computer system will have to learn and develop naturally from its environment in some capacity, and then perhaps be copied. Thus, human-like artificial intelligence will require design that centers on pattern recognition and learning. This all may sound like bad news for AI, but I think it bodes well. It means that an extremely complex, intelligent system can develop from the infancy of a few brilliant algorithms designed to achieve perpetual, adaptive self-organization through learning. We just need a few brilliant minds to discern the nascent self-organizing automata that underlie "intelligence."
posted by Clayton at 9:21 PM
0 Comments:
Post a Comment
<< Home