|
Author
|
Topic: Smallest Self Contained Unit of Intelligence
|
warren_bergerson
Member
Member # 262
|
posted 01. November 2003 07:39
Start with two generally recognized facts- 1)AI has not been nearly as successful as originally expected despite huge increases in computing capacity and huge advances in understanding the physiology of the brain and 2)the ability to solve a problem is highly dependent on how the problem is viewed or perceived. The reasonable conclusion suggested by these two facts is that we may be approaching AI from the wrong viewpoint or perspective.
Essentially all current approaches to AI are based on the assumption or perspective that the human brain or mind is a self contained unit capable of generating intelligent behavior. In other words, AI research is based on the idea that it should be possible to build or program a self contained intelligent computer because the human brain or mind is such a machine.
Is it possible that this assumption or perception is the reason that AI has not been successful? Is it possible that understanding intelligence and intelligent behavior requires recognizing that ‘intelligence’ is a larger and more complex process that can be fit in a single nervous system? Is it possible that rather than a computer, the appropriate analogy for ‘a unit capable of intelligent behavior’ is the Internet. Is it possible that a human brain is no more than a ‘smart terminal’ in an interconnected set of terminals and processors? Could the Internet analogy prove more productive in AI analysis than the brain as a self contained intelligent computer analogy?
Any comments? For a more detailed discussion of this idea, see my article at http://groups.msn.com/LifeEngineering/unitofintelligence.msnw
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 01. November 2003 15:23
The question of the smallest unit of intelligence is an interesting one. It's interesting in part because there are really two aspects to the question. First, one has to ask, what are the minimal capabilities that a system must have before we would consider it intelligent? Second, how much hardware/processing power is required and what kind of algorithms are required, if these capabilities can be manufactured at all?
I am not sure how to robustly address the first question, but I will give some examples that I find interesting to think about.
Cockroaches are tough, adaptable little creatures. They excel in finding food of all sorts (natural and manmade) in a variety of complex environments, and are good at escaping predators using a combination of visual cues and wind-current detectors. They do have the capability to learn in limited ways, but they're not really trainable as a pet. Are cockroaches intelligent?
Dogs are highly capable pack animals, able to (learn to) sense and respond to a wide variety of human emotions, complex commands, and varying conditions. For example, dogs not infrequently bring telephones to injured masters, which requires them to know that their master is injured, that the dog can't help, that other humans can help, and that the telephone is used to summon other humans. Are dogs intelligent?
Aibo is a toy robot dog produced by Sony. It is able to learn tricks, respond to human commands, and engage in sophisticated behavior. It can be trained to perform many of the simple behaviors dogs can be trained to do and cockroaches can't (e.g. commands to sit and such), so it clearly has learning capabilities superior to the cockroach, yet it is much worse at dealing with unpredictable environments than either the cockroach or a real dog. Is Aibo intelligent?
One could chauvinistically say that we only consider things intelligent if they are at the level of human intelligence. Intelligence seems a matter of degrees, even between humans (esp. considering various types of brain damage and retardation), but this can be a useful point from which to tackle the second question.
It is formally possible that intelligence is non-algorithmic, and that there is no finite algorithm of implementable size that would approximate intelligence. This would be a vitalist view of intelligence, which is a not uncommon viewpoint among philosophers (especially Christian ones).
It is also possible that we're going about intelligence all wrong, and that implementable algorithms do exist but we're badly sidetracked. I think there's a good argument to be made that many AI researchers are badly sidetracked. Much of AI research has focused on predicate logic machines, which have been found to require gazillions of rules and still have shockingly brittle responses. Yet we see that humans can perform well even when they only have fuzzy information, misconceptions, and without any formal rules at all. Predicate logic seems an unlikely candidate for building low-level intelligence. Various expert systems fall into this category also. They can produce useful, intelligent-like behavior in a narrow range of circumstances, but it is difficult to consider the entire system intelligent (at least when compared with human capabilities).
The other major approach recently has been with various artificial neural networks trained using massive data sets and performing tiny increments of internal weights after analyzing each item in the data set. These systems often end up with powerful, robust, and relatively noise-insensitive performance, but only in the area they are trained in. Given the massive quantities of data that are needed to train them, these areas are necessarily very limited, and it's often not clear how one would generalize them anyway. In contrast, humans can perform well with only a brief description of a situation, or a few trials of experience. So this system is apparently also fundamentally flawed for active online intelligence.
There are a variety of other approaches out there (Bayesian, associative matrix, explicit physical modeling of neurons, etc.), but they haven't been researched enough for us to really know how well they will pan out. As to the internet analogy...well...it's not really precise enough for me to guess.
There is, however, something that we have to keep in mind when evaluating our success in creating AI. It may be that the smallest self contained unit of intelligence is actually quite large. Our best desktop computers have less theoretical processing power than a fruit fly's brain. (Then again, our computers use a much larger fraction of their theoretical power than a fruit fly is likely to.) With timing accuracies of about 2ms and between 10^14 and 10^15 synapses, the human brain could theoretically be performing on the order of 10^18 operations per second--a hundred thousand times more than the new G5-based supercomputer at Virginia Tech that's now #4 in the world in supercomputing power. Artificial intelligence researchers never get access to anything even as powerful as the G5 cluster. So the smallest self-contained unit may simply be well beyond our capabilities, even now.
There is good reason to beleive that biological organisms' intelligence is inherent to themselves, not part of some invisible connected network. The reason is that damage, drugs, and experience had by one organism affects its performance in immediate and measurable ways; however, damage, drugs, and experience of other organisms that have no direct contact with the test organism do not appear to have impacts on performance. Of course, this doesn't rule out connection into some larger network that is beyond our ability to influence, but it seems at odds with the results of direct observation of isolated individuals, and isn't necessary to explain our computational failures given that we're vastly short of computing power to explicitly match a human (or other mammal) brain, and given that we don't really understand at the detailed mechanistic level how brains process information. We know a lot about high level systems, and about individual neurons, but how the neurons group together to provide low-level systems that group together to provide intermediate-level systems that group together to provide the high-level systems is still a mystery.
IP: Logged
|
|
nobody
Member
Member # 145
|
posted 08. November 2003 22:20
Rex asks:
quote:
Are cockroaches intelligent?
The quick and easy answer is no. However it more likely depends on your definition of intelligence, imo. There obviously is a wide range of intelligence that we can observe, even if we limit our observation to humans only.
One interesting point that has been made about the intelligence of a cockroach, I think it was pointed out by Professor Michio Kaku, is that the cockroach is smarter than any robot humans have been able to create so far. So if we claim our robots exhibit intelligence, then "yes" the cockroach is intelligent.
IP: Logged
|
|
nobody
Member
Member # 145
|
posted 08. November 2003 22:49
Hello Warren,
Your concept of the Smallest Self Contained Unit of Intelligence is certainly intriguing. You do have a certain flair for thinking outside the box, which I find a very good talent for scientists to have. Einstein had it, but it appears to me that it is quite rare today.
You ask:
quote:
Could the Internet analogy prove more productive in AI analysis than the brain as a self contained intelligent computer analogy?
I think you're onto something! This article will provide additional food for thought regarding your question:
quote:
To illustrate the brain's tremendous capacity, Sejnowski and Laughlin state that the potential bandwidth of all of the neurons in the human cortex is "comparable to the total world backbone capacity of the Internet in 2002.
Isn't that absolutely amazing?!
Here's the link:
http://www.sciencedaily.com/releases/2003/10/031002053904.htm
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 10. November 2003 08:30
Nobody,
The topic introduced here was intended to relate primarily to the development of formal criteria used in developing AI simulations. If intelligence is viewed or approached as a unit larger than the human head, then the requirements for an AI system must include recognition of the role of input from other parts of the intelligent process (like input to an Internet terminal from other components of the Internet). The size of intelligence issues is also intended to focus on the role of information transmitted from the past(front loading).
The efforts to build systems capable of simulating ‘high level creative human intelligence’ have not been successful. The larger than the head view of human intelligence suggests that these failures may not be due to inadequate computer power or lack of appropriate processing algorithms. The AI failures may in large part due to the fact that the requirements being used in AI do not properly reflect the dispersed processing, input, and front loading associated with human intelligence.
Quote: which I find a very good talent for scientists to have. Einstein had it, but it appears to me that it is quite rare today
The portion of humans who are creative versus conformists seems to be fairly constant as is the amount of creativity. The factor that appears to vary from time to time and sub-culture to sub-culture is the acceptance or non-acceptance of practices to suppress creative or unconventional ideas. The message board on this site- http://groups.msn.com/LifeEngineering/messages.msnw has a discussion of a proposed "Corruptive Practices Project" designed to analyze this topic. You may find it interesting
[ 10. November 2003, 08:32: Message edited by: warren_bergerson ]
IP: Logged
|
|
nobody
Member
Member # 145
|
posted 10. November 2003 12:44
Good morning Warren,
I must respectfully differ with you here:
quote:
If intelligence is viewed or approached as a unit larger than the human head
If you are looking for the smallest unit of intelligence, you need to define it in some way that I cannot yet understand. But however you define I think it is clear that a human has many many units of intelligence. A cockroach would be far down the scale, much closer to having one unit of intelligence, imo.
I assume you have considered the variation of intelligence between humans. We try to quantify that with IQ tests, but I think you are searching for something other than IQ.
Anyway, we have the term "ohm" still being used in connection with electricity so, if you can define your proposed unit of intelligence, I propose that it must logically be named the "bergerson".
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 11. November 2003 05:58
Nobody,
Viewing intelligence as ‘larger than a human head’ means viewing intelligence as being made up of many, many different types of sub-processes, not all of which are located inside a single human head. The ‘mega-intelligence’ point of view characterizes intelligence as a forest or mountain range rather than as trees or rocks.
Intelligent behavior is the product of information processing. The purpose of the mega-intelligence perspective is recognition of the fact that a portion of the information processing responsible for a particular manifestation of ‘intelligent behavior’ may have occurred at a time and a place other than the immediate vicinity of manifestation.
Example, a person writes a book. The suggestion that the intelligence or information processing responsible for writing a book all occurred in the brain or mind of the writer is misleading or potentially misleading in two respects. First, it does not recognize the role of social interaction with other people (and with the physical environment). Second it does not recognize the role of the processing or pre-processing that may have occurred over millions or billions of years in the past.
The question of where information processing occurs is particularly important in understanding the unique characteristics of human intelligence. Information processing in humans, intelligence in humans, is highly dependent on information processed or pre-processed by others. Failure to recognize the role of externally generated information is probably the main reason AI has been unsuccessful at simulating ‘high level human cognitive processes’. As an example, high level human decision making involves very complex information processing. This complexity is readily recognized by those attempting to simulate this type of behavior. They, however, incorrectly assume that all the complex processing occurs within a single human head. Once you recognize the role of social communications, the processing performed in a single human head becomes relatively easy to understand and simulate. In this example, perspective is the key to solving the ‘how is it done problem’.
The question of ‘when’ information processing occurs is actually far more important to understanding intelligence. Intelligence is clearly progressive. Intelligent processing or behavior in the past clearly forms (or can form)the basis or ground work for more ‘advanced’ intelligent behavior in the future. Recognition of the cumulative or progressive nature of intelligence and intelligent behavior is arguably the most important feature of the mega-intelligence perspective. The traditional naīve viewpoint is that intelligence is a phenomena which creates intelligent behavior ‘from scratch’ in a very short period of time. The mega-intelligence perspective, by contrast, suggest that intelligence is a process of set of processes working over very long periods of time which finally cumulate in observed occurrences of intelligent behavior.
It is worth repeating that the issue here is ‘perspective’ not fact. How we view data influences how we interpret it. If we view the human mind as a self contained system, then we tend to view input from outside sources and input/processing from the past as irrelevant. Such a priori dismissal of factors may, and I suggest has, contributed to the failure to develop AI systems.
It should also be noted that understanding intelligence and the scientific analysis of intelligence involves the use of many different perspectives. The all life forms or mega-intelligence perspective should be viewed not as ‘the’ correct viewpoint, but one of many different useful viewpoints.
The mega-intelligence perspective does not prevent us from quantifying the intelligence of humans, cockroaches and single cells. The perspective simply means that in quantifying the intelligence of sub-units we are only quantifying some limited aspect or feature of intelligence. We can quantify the lift produced by a wing, or the thrust of a jet engine even though we an ‘airplane as entity larger than a wing or engine.
IP: Logged
|
|
nobody
Member
Member # 145
|
posted 11. November 2003 12:34
Okay....
So how does this perspective help move AI forward? I'm not seeing a practical application yet.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 11. November 2003 13:42
Some people in AI have been working on what is apparently the "mega-intelligence" issue for some years. The notion that broad background knowledge is essential for a working AI to be non-brittle isn't new. Doug Lenat started the "Cyc" project in the late 1980's: quote:
OpenCyc is the open source version of the Cyc technology, the world's largest and most complete general knowledge base and commonsense reasoning engine. Cycorp, the builders of Cyc, have set up an independent organization, OpenCyc.org, to disseminate and administer OpenCyc, and have committed to a pipeline through which all current and future Cyc technology will flow into ResearchCyc (available for R&D in academia and industry) and then OpenCyc.
RBH
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 11. November 2003 15:17
It is highly doubtful in any outlook or perspective is ever entirely new and I certainly did not suggest I originated the mega intelligence idea. I am not familiar with Cyc specifically, but my understanding was these were primarily search engines like google allowing individuals to efficiently search large data bases.
Looking at something from a new perspective obviously does not automatically produce a solution. Looking at a moldy petri dish as ‘interesting’ rather than ‘a messed up experiment’ was only a starting point.
The mega-intelligence perspective is useful in that it emphasizes the role of dispersed or distributed information processing in producing intelligent behavior. Dispersed processing means that intelligent behavior is or can be the result of 1)information processing performed at a physical distance and 2)information process performed long in the past.
One ‘practical’ use of this concept is in developing formal mathematical or engineering designs for reverse engineering intelligent systems or for designing AI systems. For an ongoing discussion of the formal design application see the automated learning topic at http://www.generation5.org/forums/forum.asp?FORUM_ID=2 .
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 11. November 2003 16:00
warren_bergerson wrote quote:
For an ongoing discussion of the formal design application see the automated learning topic at (this URL).
Do you mean the "automatic learning" thread? There is no "automated learning" thread.
RBH
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 11. November 2003 20:17
The individual units--a human, for example--has to pick up all that external intelligence from somewhere, though. And AI researchers certainly are aware of this. Every serious AI project that I know about assumes that machine learning systems will have to learn a great deal about their environment before becoming functional, just as humans do. The method of learning is problematic; nobody has come up with systems that mimic human capabilities in that area.
Machine learning is not synonymous with AI. Because of the immense difficulty in actually getting anything worthy of the name AI, many people concentrate on machine learning in much more limited areas. I agree that they often ignore the structure of knowledge and complicated distributed source of events that go into intelligence, but my sense is that historically this is a reaction to being successful in limited areas when ignoring most of the complexity, and failing miserably in all areas when not.
Perhaps it's about time to start going back the other way, though. Machine learning has progressed a lot in the last two decades. (Not nearly enough, I don't think, but I would be happy to be mistaken.)
[ 12. November 2003, 12:21: Message edited by: Rex Kerr ]
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 12. November 2003 11:56
As Rex’s comments suggest, the discussion of AI viewpoints or perspectives leads quite naturally to a discussion of past AI successes or failures. Past failures may be due to looking at an issue from an unproductive perspective. Past successes mean that an issue was addressed from a productive perspective.
There are those who would argue that AI results to date have been disappointing, at least with respect to the key goal of simulating the types of high level intelligent behavior associated with humans. This interpretation of AI results may not be accurate. As an alternative interpretation, I suggest, that while no one has yet produced a successful simulation of high level intelligent human behavior, the major technical/conceptual problems associated with building such a simulation have been resolved (although the existence of these solutions is not always recognized).
To fully appreciate the progress made to date, it is useful to view the development of AI in terms of an engineering process. In simple terms, an engineering process consists of an abstract design phase and an implementation phase. To properly understand the current state of AI, I suggest it is useful to ask if AI failures were due to design phase failures or to implementation stage failures. We can then look at the AI successes to see what they did differently to produce successes. Finally, we go back to the failures and determine if the techniques and concepts that produced successes could be used to address the problems which caused the AI failures.
In starting this type of review, we need to recognize that much of the development of AI systems is described as being based on an informal design and development process (much the way very early airplanes were developed using informal design processes). However, it appears that individuals actually engaged in the successful design and development of complex computer systems use techniques, concepts and processes that are very similar to formal engineering design and development processes.
FAILURES IN THE IMPLEMENTATION STAGE
AI systems involve input processes, output processes, and logical processing of information. Input and output devices involve a variety of complex technical issues, but for the most part there are well established techniques for addressing such issues. AI failures in the implementation process thus come down to failures in information processing. Since there are an essentially unlimited number of different ways of performing information processing given an appropriate design, failures at the implementation stage come down to issues of information processing capacity.
As anyone who has done complex design and programming knows, there are lots of highly effective techniques for analyzing and addressing capacity issues. Again to greatly simplify, if you have a sound design, capacity problems are addressed by 1)breaking processing into pieces and performing different pieces at different times and places (distributed processing) or 2)going back and developing a more efficient design.
It is therefore reasonable to conclude that if we had a sound design for human AI simulation, we could solve the processing capacity problem.
FAILURE AT THE DESIGN STAGE
AI failures, it would appear, are due to failures at the design stage. This may involve failure to develop a sound initial design or failure do adequately improve the efficiency of an initial design. Based on only a rough review, it appears that failure to develop an initial design are of two primary types- 1)never got to the formal design stage (never got beyond the informal ‘user specs’ stage) and 2)designs built around a concept such as a GA or a neural network(efforts to build complex systems around one or two components). There obviously may be other types of design stage failures.
It is interesting to note that in the instances where developers succeeded in building an initial workable prototype, as with voice recognition, it has almost always proved possible to go back and improve on the design and performance.
LOOKING AT DESIGN STAGE SUCCESSES
There are a wide range of AI successes including robotic control of motion, and voice and visual recognition processes. Again, this is based on a rough review, but it would appear that all the successful simulations shared the following design stage characteristics:
1. All involve precise definitions of the behavior and processes being simulated.
2. All involve goal-oriented or purposeful behavior.
3. All are progressive or improveable.
It might also be added that as far as I am aware, in no instance did a design process which had a precise definition of the phenomena being simulated run into a process or phenomena which could not be modeled or simulated.
CONCLUSION
Recognizing that opinions on the subject can vary, my personal interpretation of AI results to date is that most of the technical roadblocks to building a system capable of simulating human intelligence have already been resolved. It should, IMO, be possible to design, and to build at least a prototype simulation if we :
1. Adhere to rigorous engineering design standards and principles. It is particularly important to stick to the standards relating to defining variables and starting point states or conditions.
2. Recognize that intelligent behavior is generated by goal-oriented or teleological processes.
3. Recognize the role of distributed processing.
As noted above, results to date suggest that if you start with a sound initial abstract or engineering design, and if you can build a simple prototype of that design, then it is reasonable to expect that the prototype can be enhanced to eventually match or exceed human performance.
IP: Logged
|
|
nobody
Member
Member # 145
|
posted 13. November 2003 14:27
quote:
it is reasonable to expect that the prototype can be enhanced to eventually match or exceed human performance.
Okay Warren,
I'm a bit of a skeptic, so let me go on record and state that I do not expect this to happen within my lifetime. If I'm lucky, that would be roughly another quarter of a century.
IP: Logged
|
|
Steve Petermann
Member
Member # 884
|
posted 14. November 2003 11:31
Interesting phrase, "smallest self-contained unit of intelligence". Seems to me this question strikes at the heart of what intelligence is and how it works.
The first issue would be whether intelligence can be "self-contained". While philosophers have perennially debated the question of a priori knowledge, current biology and psychology would seem to suggest that while organisms may represent a container of sorts, only those that can interact with their environment and move beyond a priori capabilities have any chance of displaying intelligence. As Donald Davidson claims, even things like language and with it thought are only created because of external relationships.
This would also suggest that some form of a mega-intelligence perspective is an essential component to understanding intelligence. If this is true then the idea of "self contained intelligence" is a non-starter.
To flesh out the idea that intelligence requires interaction with the environment including all its entailments throughout history, I would offer that the smallest unit of intelligence would contain these elements: input and output with an environment, a engine for persistent representation and a correlative engine. I base the need for these components on the idea that the absolute essential component of intelligence is abstraction.
I believe examples from current neurobiology can support this. The human brain has the ability to create representations of its environment in the form of various neural structures and trajectories. Initially these neural representations are concrete representations of the environment. However, if the brain was only able to create concrete representations it would be hard to see any benefit from that mechanism. To be of use these concrete representations have to be correlated, become relational. It is only in correlation that there can be a move from instinctive hard coded behavior. What the correlation engine in the brain does, however, is create abstractions, new neural structures and trajectories that are not bound by the concrete. An analogy could be an electronic circuit. If a circuit is hard wired then only stereotypical, mechanistic behavior can ensue. However, it there is also an internal mechanism to "rewire" the circuit then there can be novelty and "learning". This seems to be what happens in the human brain. Not only does the brain have the ability to pick up sensory information and represent it, it also has the ability to create correlations. Physically these occur through new connectivity between neural structures. This is accomplished by growing dendrites. When this new connectivity occurs something novel is created, an abstraction that includes concrete representations but also transcends them. Interestingly enough, this also provides for the ability to embrace external abstractions and correlate then to existing structures. The greater and more efficient this correlation and abstraction process is, the more intelligence ensues. Now it is possible for strictly internal processing(reflection) to create new abstractions, but that obviously has a limited scope. However interaction with external abstractions via culture and experience offer an almost unlimited potential for expanded and more efficient intelligence.
If this seems a reasonable scheme, it means that wherever systems with: input, output, representational and correlative engines are available, there can be intelligence. Certainly higher level organisms have these components and possibly even the lowly cockroach. Although bacteria and virus can "adapt" to there environment, they would not be considered intelligent. This also means that for AI schemes to show advanced intelligence, they would need to have interactions outside themselves and draw correlations from those interactions with their current internal representations.
IP: Logged
|
|
|
Printer-friendly view of this topic