Artifical Intelligence (AI) are there for many years, but not as a comprehensive successful solution yet, due to the lack of measurement of intelligence and dynamics of intelligence.
Measurement plays critical roles in the development of dynamics: clock, telescope, microscope, etc. We even do not have a good way to measure intelligence now.
The concept Computational Complexity is used to measure computation, not intelligence. Current supercomputers already have comparable computing power and memory as human brains, but they are not even close to human intelligence now.
This is a clear indication: computation is different from intelligence, so Turing Machine and Computational Complexity Theory are not a right theoretic framework for intelligence and AI.
Turing Machine or equivalences, are actually highly influenced by deterministic models. Back to the dynamics for the real world, there are quantum theories, etc., which suggest the possibility of nondeterministism.
Some scientists tried to build the human mental models based on quantum theories. However, the difference between the time scale of neuron firing and excitations in microtubules and the decoherence time tells us there are some links in the middle still missing.
That is the reason I propose this three-level model: Life, Psychology, and Intelligence, to fit in the gap. Psychology is a subdomain of life, built on top of the other parts of life. Intelligence is a subdomain of both life and psychology, and built on top of the other parts of both life and psychology.
Ilya Prigogine's theory of systems far from equilibrium is a good foundation for life phenomena. My goal is to further study and accumulate the knowledge and models of psychology based on Brussels-Austin group and Ilya Prigogine's theories. Once concrete enough foundation has been constructed for this middleware, we coud combine the quantum theories and intelligence models.
So far, most of AI researches are based on methods of equilibrium. These approaches are difficult to make progresses in many areas. Probably people should try to re-focus AI researches on methods of far from equilibrium, rather than only on those of equilibrium.
There are some discussions of ontological aspect and epistemic aspect in Brussels-Austin approach. In real world, due to measurability, there could be several possibilities:
1) Measurable Determinism, determined by measurable factors: environment and internal states, etc.
2) Unmeasurable Determinism, determined by some factors which are not measurable.
3) Nondeterminism. With this model, there could be somethings such as free wills, etc.
Systems far from equilibrium, combined with the possibilities of measurable or unmeasurable, could be the way to illustrate the complexity of psychology and intelligence.
Based on my three-level model, the psychological concepts such as 'will' or 'free will' could be re-studied based on systems far from equilibrium. New concepts and models of intelliegnce could be proposed based on further studies.
Since life is a typical type of systems far from equilibrium, very different from other systems, Brussels-Austin group approach even could gain some hints from the studies based on my propsoal, to re-energize and push forward their research.
In this proposal, there are seven main points, if no one proposed them before, are my new contributions:
1) Propose the concept of Dynamics of Intelligence
2) AI fails as a comprehensive solution so far, due to the lack of measurement of intelligence and dynamics of intelligence.
3) Computation is different from intelligence. Computational complexity is used to measure computation, not intelligence. Turing Machine is not a right theoretic framework for artificial intelligence.
4) The difference between the time scale of neuron firing and excitations in microtubules and the decoherence time, does not exclude the possiblility to build psychology and intelligence models on quantum theories. Just people need build the missing middleware between them first.
5) The theories of far from equilibrium could be used as the foundation of the middleware mentioned in 3).
6) AI researches better to re-focus on methods of far from equilibrium, rather than only on those of equilibrium.
7) Brussels-Austin approach could use life phenomena as good study targets to gain hints, re-energize their efforts, and push forward the researches.
8) A three-level model for intelligence: Life, Psychology, and Intelliegnce. Psychology depends on the other parts of life. Intelligence depends on the other parts of both life and psychology.
If someone already proposed some or all of these ideas before, please let me know. I would really appreciate.
Friday, August 19, 2011
Saturday, August 6, 2011
The Proposal for A Three-level Intelligence Model
I wrote three articles for my proposal of a new intelligence model, a pure research proposal.
In The Things AI and Robots Can Do Well and Not Well So Far, I summarizes the current status of AI and robotics, to justify the need to develop more advanced AI theories and methodologies.
In Life, Psychology, and Intelligence, I introduced a three-level model of intelligence, based on quantum theories and systems of far from equilibrium. Measurability or unmeasurability highly affect knowledge models, and life is a type of systems far from equilibrium. These are critical factors for a generic intelligence model.
In Reductionism and Whole/Parts, I described the philosophical reasoning to build this model on low level theories: quantum and far from equilibrium systems. This model is not pure reductionism as explained there.
This model assumes the existence of the physical world. Although unmeasurability is mentioned, it doesn't take agnostic position. Because things could be measured up to certain degrees, and we don't know where the up limit is yet even if it DOES exist.
In The Things AI and Robots Can Do Well and Not Well So Far, I summarizes the current status of AI and robotics, to justify the need to develop more advanced AI theories and methodologies.
In Life, Psychology, and Intelligence, I introduced a three-level model of intelligence, based on quantum theories and systems of far from equilibrium. Measurability or unmeasurability highly affect knowledge models, and life is a type of systems far from equilibrium. These are critical factors for a generic intelligence model.
In Reductionism and Whole/Parts, I described the philosophical reasoning to build this model on low level theories: quantum and far from equilibrium systems. This model is not pure reductionism as explained there.
This model assumes the existence of the physical world. Although unmeasurability is mentioned, it doesn't take agnostic position. Because things could be measured up to certain degrees, and we don't know where the up limit is yet even if it DOES exist.
The Difference between Life and Machine
Machines are very useful. However, the tasks they could do are constrained by many factors and grow slowly now. To explore the potential of AI to sustain the growth, people need understand the differences between life and machines
.
Mechanical clocks and watches with extremely high precision and sophistication appeared long, long ago before computers, as well as those old automatons. The steam engines and centrifugal governors back in 18th century can generate power much better than humans. However, they do not have AI. They are only machines.
It is obvious there are still huge differences between life and current machines.
The difficult part is what cause the differences and how much could be reduced. Similiar debates on this topic were repeated again and again for decades without much progress.
This is a clear indication we do not have a proper theoretic framework for AI yet to anwser this question. We even cannot define the most basic concept for AI:the intelligence itself.
Doing some work better than humans is not an evidence of intelligence. A great machine does not necessarily have great AI, which is still very limited in its applications.
There is a concept: strong AI, which means: "the intelligence of a machine that can successfully perform any intellectual task that a human being can" (from wikipedia). Obvious, there is not strong AI so far.
We even do not have a way to evaluate the current status, to measure how far away an AI systems are from strong AI, which may causes illusions to many people.
We have a concept: weak AI, which does not "match or exceed the capabilities of human beings, as opposed to strong AI" (from wikipedia). However weak AI is only a vague phrase, which does not measure how strong or how weak an AI system is.
This is another indication people do not have a proper theoretic framework for AI so far.
To have a good estimation how strong an AI system is, we could redefine strong AI as strong human AI, and break down the big category of weak AI into those such as:
1) strong ape AI, match or exceed apes' capability in all intellectual aspects
2) strong monkey AI, match or exceed monkeys' capability in all intellectual aspects
3) strong cat AI, match or exceed cats' capability in all intellectual aspects
4) strong bee AI, match or exceed bees' capability in all intellectual aspects
5) strong fish AI, match or exceed fishes' capability in all intellectual aspects
etc...
People could add more as they like. When using these fine-grained categories, at least people can have clear sense what are the differences between AI systems and a specific type of life.
Looks we don't have strong cat AI yet. Technical singularity is still far away.
Strong human AI is a superset of both life and machine. We don't know if it is achievable or not yet. It remains as the last boundary to cross between life and machine. I propose a three-level model of Life, Psychology and Intelligence. Hope further research on this model could illustrate how close machines could be made to life and humans.
.
Mechanical clocks and watches with extremely high precision and sophistication appeared long, long ago before computers, as well as those old automatons. The steam engines and centrifugal governors back in 18th century can generate power much better than humans. However, they do not have AI. They are only machines.
It is obvious there are still huge differences between life and current machines.
The difficult part is what cause the differences and how much could be reduced. Similiar debates on this topic were repeated again and again for decades without much progress.
This is a clear indication we do not have a proper theoretic framework for AI yet to anwser this question. We even cannot define the most basic concept for AI:the intelligence itself.
Doing some work better than humans is not an evidence of intelligence. A great machine does not necessarily have great AI, which is still very limited in its applications.
There is a concept: strong AI, which means: "the intelligence of a machine that can successfully perform any intellectual task that a human being can" (from wikipedia). Obvious, there is not strong AI so far.
We even do not have a way to evaluate the current status, to measure how far away an AI systems are from strong AI, which may causes illusions to many people.
We have a concept: weak AI, which does not "match or exceed the capabilities of human beings, as opposed to strong AI" (from wikipedia). However weak AI is only a vague phrase, which does not measure how strong or how weak an AI system is.
This is another indication people do not have a proper theoretic framework for AI so far.
To have a good estimation how strong an AI system is, we could redefine strong AI as strong human AI, and break down the big category of weak AI into those such as:
1) strong ape AI, match or exceed apes' capability in all intellectual aspects
2) strong monkey AI, match or exceed monkeys' capability in all intellectual aspects
3) strong cat AI, match or exceed cats' capability in all intellectual aspects
4) strong bee AI, match or exceed bees' capability in all intellectual aspects
5) strong fish AI, match or exceed fishes' capability in all intellectual aspects
etc...
People could add more as they like. When using these fine-grained categories, at least people can have clear sense what are the differences between AI systems and a specific type of life.
Looks we don't have strong cat AI yet. Technical singularity is still far away.
Strong human AI is a superset of both life and machine. We don't know if it is achievable or not yet. It remains as the last boundary to cross between life and machine. I propose a three-level model of Life, Psychology and Intelligence. Hope further research on this model could illustrate how close machines could be made to life and humans.
Wednesday, July 20, 2011
Reductionism and Whole/Parts
In my previous post, I proposed a three-level model of intelligence: Life, Psychology, and Intelligence, which is based on the theories of far from equilibrium which again could be based on quantum theories. So total 5-level models, including life and non-life. This article explains some philosophical reasoning behind this model.
These models sounds like reductionism, but not exactly.
Let's first consider the phrase "the whole is greater than the sum of its parts", which has been talked by many, many people and many, many times.
In reality, the highly abstracted "whole" could be much less than the sum of its parts, if people ignore too many details and don't have enough knowledge of parts.
So "the whole is greater than the sum of its parts" is only true when you have the enough knowledge of parts. Usually it is not the case. People should be aware of the constraints of those sayings which sounds correct.
However, in practice, many people still follow such type of thinking, not exactly, but actually in a loosed form, consciously or not consciously, "the whole could provide additional information to the sum of its parts". Here the constraint of "enough knowledge of parts" is removed. But pay attention to "could". You take the risk for your own by following this thinking. It only COULD be true.
The constraint tells us, people need pay attention to the details of parts, especially when things are not done well. By coming back to fundamentals, people could enrich their understanding, adjust their approaches.
However, only low-level details are not good enough. By the loosed version of "the whole could provide additional information to the sum of its parts", people could establish high-level principles, which may be not obvious in low-level regular phenomina, to understand the whole better. Brussels-Austin approach could gain hints from the studies of life, psychology and intelligence, to push their researches further.
So although "the whole is greater than the sum of its parts" is not always true, this three-level model of intelligence is not pure reductionism.
These models sounds like reductionism, but not exactly.
Let's first consider the phrase "the whole is greater than the sum of its parts", which has been talked by many, many people and many, many times.
In reality, the highly abstracted "whole" could be much less than the sum of its parts, if people ignore too many details and don't have enough knowledge of parts.
So "the whole is greater than the sum of its parts" is only true when you have the enough knowledge of parts. Usually it is not the case. People should be aware of the constraints of those sayings which sounds correct.
However, in practice, many people still follow such type of thinking, not exactly, but actually in a loosed form, consciously or not consciously, "the whole could provide additional information to the sum of its parts". Here the constraint of "enough knowledge of parts" is removed. But pay attention to "could". You take the risk for your own by following this thinking. It only COULD be true.
The constraint tells us, people need pay attention to the details of parts, especially when things are not done well. By coming back to fundamentals, people could enrich their understanding, adjust their approaches.
However, only low-level details are not good enough. By the loosed version of "the whole could provide additional information to the sum of its parts", people could establish high-level principles, which may be not obvious in low-level regular phenomina, to understand the whole better. Brussels-Austin approach could gain hints from the studies of life, psychology and intelligence, to push their researches further.
So although "the whole is greater than the sum of its parts" is not always true, this three-level model of intelligence is not pure reductionism.
Things AI/Robots Can Do Well and Not Well So Far
1) AI can play chess quite well. Chess is played on small 8x8 boards with strict rules, a task extremely fitting to AI. However, chess softwares are continuely being improved by humans.
Even for chess, I don't know for a fixed chess software without any farther tuning from humans, how long it can keep winning against top human players. Since humans can learn better than AI so far.
2) AI can play GO well, but not so well as chess, due to combinatorial explosion. GO is played on 19x19 boards, also with strict rules.
3) Auto car pilots. It runs on roads built for cars and with highly-regulated traffic rules. Only a few cases are reported so far. Don't know how well these systems can run in large scale and randomly picked road situations.
A reasonable guess is: potentially auto car pilots could do well in standard car racing competition if not already, which runs on routes regulated better than usual roads, easier for safety. Although improving performance might be more difficult initially than running at normal speeds, safety could be more difficult to solve eventually.
4) Robots' capabilities are highly constrained by AI.
They can play Rubiks Cube extremely well, since there are fixed algorithms. They can play music instruments well, with fixed actions. They can even play pingpong with humans, which is on simple platform with simple rules. Hopefully, robots could catch up with or even exceed humans soon with pingpong.
They can walk. With two legs, they can walk on flat places, or stairs with regular shapes, but not good at mountain surfaces yet. With four legs, they could walk on rough terrains, but may not on very steep hills or rocky mountains as deers and other animals can, not to mention to find a route by themselves there. With six legs, they can even climb on coarse walls or trees, but not seen on glass or steel walls (by me) yet, probably would in near future since glass or steel wall surface is well-defined.
They could emulate simple creatures like fishes, moths, somehow snakes, but not even behave like chicken, rats, cats, dogs, etc., yet so far.
So even to achieve goals of weak AI, robots still need much more advanced AI theories and methodologies.
There is one category: industrial robots, which already play significant roles in reality. They are designed for well-defined tasks. However, there are lessons could be taken from the Saturn project in 1980s, if people want to make big breakthrough.
5) Handwriting recognization, voice recognization, image and video-processing, etc. These can do well in certain contexts and with specific targets to recognize, such as moving objects. AI could recognize what people write better if it already knows what type of things are written: digitals, English letters, Chinese characters, mathematical signs, etc. Or for voice: what languages people are speaking, or they are singing a song, or just playing Kouji, etc.
......
People could add more to this list. However the list would still not be long. In most situations, computers only can aid and enhance humans, not replace them.
AI cannot do most things humans can. So IMHO, AI is still in poineering stage. Only a few things could be implemented in software so far. More research is needed, including much philosophical thinking.
Don't limit youself to software implementation in AI researches. Most of them cannot be implemented so far.
Even for chess, I don't know for a fixed chess software without any farther tuning from humans, how long it can keep winning against top human players. Since humans can learn better than AI so far.
2) AI can play GO well, but not so well as chess, due to combinatorial explosion. GO is played on 19x19 boards, also with strict rules.
3) Auto car pilots. It runs on roads built for cars and with highly-regulated traffic rules. Only a few cases are reported so far. Don't know how well these systems can run in large scale and randomly picked road situations.
A reasonable guess is: potentially auto car pilots could do well in standard car racing competition if not already, which runs on routes regulated better than usual roads, easier for safety. Although improving performance might be more difficult initially than running at normal speeds, safety could be more difficult to solve eventually.
4) Robots' capabilities are highly constrained by AI.
They can play Rubiks Cube extremely well, since there are fixed algorithms. They can play music instruments well, with fixed actions. They can even play pingpong with humans, which is on simple platform with simple rules. Hopefully, robots could catch up with or even exceed humans soon with pingpong.
They can walk. With two legs, they can walk on flat places, or stairs with regular shapes, but not good at mountain surfaces yet. With four legs, they could walk on rough terrains, but may not on very steep hills or rocky mountains as deers and other animals can, not to mention to find a route by themselves there. With six legs, they can even climb on coarse walls or trees, but not seen on glass or steel walls (by me) yet, probably would in near future since glass or steel wall surface is well-defined.
They could emulate simple creatures like fishes, moths, somehow snakes, but not even behave like chicken, rats, cats, dogs, etc., yet so far.
So even to achieve goals of weak AI, robots still need much more advanced AI theories and methodologies.
There is one category: industrial robots, which already play significant roles in reality. They are designed for well-defined tasks. However, there are lessons could be taken from the Saturn project in 1980s, if people want to make big breakthrough.
5) Handwriting recognization, voice recognization, image and video-processing, etc. These can do well in certain contexts and with specific targets to recognize, such as moving objects. AI could recognize what people write better if it already knows what type of things are written: digitals, English letters, Chinese characters, mathematical signs, etc. Or for voice: what languages people are speaking, or they are singing a song, or just playing Kouji, etc.
......
People could add more to this list. However the list would still not be long. In most situations, computers only can aid and enhance humans, not replace them.
AI cannot do most things humans can. So IMHO, AI is still in poineering stage. Only a few things could be implemented in software so far. More research is needed, including much philosophical thinking.
Don't limit youself to software implementation in AI researches. Most of them cannot be implemented so far.
Life, Psychology, and Intelligence: A Three-Level Model of Intelligence
Studies of Artifical Intelligence (AI) are still very difficult in many areas, such as in Robotics, etc. When things are not going well enough, people have to come back to fundamentals.
Traditionally, many of AI researches are based on Turing Machine or equivalences, which are highly influenced by deterministic models.
However, for the real world, there are quantum theories, etc., which suggest the possibility of nondeterministism.
Some scientists tried to build the human mental models based on quantum theories. However, the difference between the time scale of neuron firing and excitations in microtubules and the decoherence time tells us there are some links in the middle still missing.
That is the reason I propose this three-level model: Life, Psychology, and Intelligence, to fit in the gap.
Psychology is a subdomain of life, built on top of the other parts of life. Intelligence is a subdomain of both life and psychology, and built on top of the other parts of both life and psychology.
Ilya Prigogine's theory of systems far from equilibrium is a good foundation for life phenomena. My goal is to further study and accumulate the knowledge and models in psychology based on Brussels-Austin group and Ilya Prigogine's theories. Once concrete enough foundation has been constructed for this middleware, we could combine the quantum theories and intelligence models.
So far, most of AI researches are based on methods of equilibrium. These approaches are difficult to make progresses in many areas. Probably people should try to re-focus AI researches on methods of far from equilibrium, rather than only on those of equilibrium.
There are some discussions of ontological aspect and epistemic aspect in Brussels-Austin approach. In real world, due to measurability, there could be several possibilities:
1) Measurable Determinism, determined by measurable factors: environment and internal states, etc.
2) Unmeasurable Determinism, determined by some factors which are not measurable.
3) Nondeterminism. With this model, there could be somethings such as free wills, etc.
Systems far from equilibrium, combined with the possibilities of measurable or unmeasurable, could be the way to illustrate the complexity of psychology and intelligence.
Based on my three-level model, the psychological concepts such as 'will' or 'free will' could be re-studied based on systems far from equilibrium. New concepts and models of intelliegnce could be proposed based on further studies.
Since life is a typical type of systems far from equilibrium, very different from other systems, Brussels-Austin group could even gain some hints from the studies based on my propsoal, to re-energize and push forward their researches.
In this proposal, there are five main points which, if no one proposed before, are my new contributions:
1) The difference between the time scale of neuron firing and excitations in microtubules and the decoherence time, does not exclude the possiblility to build psychology and intelligence models on quantum theories. Just people need build the missing middleware between them first.
2) The theories of far from equilibrium could be used as the foundation of the middleware mentioned in 1).
3) AI researches better to re-focus on methods of far from equilibrium, rather than only on those of equilibrium.
4) Brussels-Austin approach could use life phenomena as good study targets to gain hints, re-energize their efforts, and push forward the researches.
5) A three-level model for intelligence: Life, Psychology, and Intelliegnce. Psychology depends on the other parts of life. Intelligence depends on the other parts of both life and psychology.
If someone already proposed some or all of these ideas before, please let me know. I would really appreciate.
Traditionally, many of AI researches are based on Turing Machine or equivalences, which are highly influenced by deterministic models.
However, for the real world, there are quantum theories, etc., which suggest the possibility of nondeterministism.
Some scientists tried to build the human mental models based on quantum theories. However, the difference between the time scale of neuron firing and excitations in microtubules and the decoherence time tells us there are some links in the middle still missing.
That is the reason I propose this three-level model: Life, Psychology, and Intelligence, to fit in the gap.
Psychology is a subdomain of life, built on top of the other parts of life. Intelligence is a subdomain of both life and psychology, and built on top of the other parts of both life and psychology.
Ilya Prigogine's theory of systems far from equilibrium is a good foundation for life phenomena. My goal is to further study and accumulate the knowledge and models in psychology based on Brussels-Austin group and Ilya Prigogine's theories. Once concrete enough foundation has been constructed for this middleware, we could combine the quantum theories and intelligence models.
So far, most of AI researches are based on methods of equilibrium. These approaches are difficult to make progresses in many areas. Probably people should try to re-focus AI researches on methods of far from equilibrium, rather than only on those of equilibrium.
There are some discussions of ontological aspect and epistemic aspect in Brussels-Austin approach. In real world, due to measurability, there could be several possibilities:
1) Measurable Determinism, determined by measurable factors: environment and internal states, etc.
2) Unmeasurable Determinism, determined by some factors which are not measurable.
3) Nondeterminism. With this model, there could be somethings such as free wills, etc.
Systems far from equilibrium, combined with the possibilities of measurable or unmeasurable, could be the way to illustrate the complexity of psychology and intelligence.
Based on my three-level model, the psychological concepts such as 'will' or 'free will' could be re-studied based on systems far from equilibrium. New concepts and models of intelliegnce could be proposed based on further studies.
Since life is a typical type of systems far from equilibrium, very different from other systems, Brussels-Austin group could even gain some hints from the studies based on my propsoal, to re-energize and push forward their researches.
In this proposal, there are five main points which, if no one proposed before, are my new contributions:
1) The difference between the time scale of neuron firing and excitations in microtubules and the decoherence time, does not exclude the possiblility to build psychology and intelligence models on quantum theories. Just people need build the missing middleware between them first.
2) The theories of far from equilibrium could be used as the foundation of the middleware mentioned in 1).
3) AI researches better to re-focus on methods of far from equilibrium, rather than only on those of equilibrium.
4) Brussels-Austin approach could use life phenomena as good study targets to gain hints, re-energize their efforts, and push forward the researches.
5) A three-level model for intelligence: Life, Psychology, and Intelliegnce. Psychology depends on the other parts of life. Intelligence depends on the other parts of both life and psychology.
If someone already proposed some or all of these ideas before, please let me know. I would really appreciate.
Subscribe to:
Posts (Atom)