|
Author
|
Topic: A Definition of Intelligence
|
Rex Kerr
Member
Member # 632
|
posted 30. December 2003 13:34
Warren wrote
quote:
A behavior, process or system is characterized as intelligent if and only if the behavior, process, or system is dependent on, associated with, and a product of a complex and extended preliminary processing called ‘learning efficient problem solving’ (LEPS).
but the best definition of LEPS is (from the MSN groups link)
quote:
[A system is LEPS if] 1)the behavior, process, or system is associated with a life form or biological information processing and 2)the behavior, process or system does not occur in nature except associated with a life form or biological information processing
(The mathematical definition depends on the idea of "intelligent behavior", which isn't very instructive if one is wondering what is intelligent and what is not.)
The problem with this definition is twofold. First, it is too broad, or simply includes everything associated with living beings by fiat. For instance, production of heat from sugar via ATP is only associated with living beings, but it seems unhelpful to call this "LEPS". Second, it is unclear why intelligence must be limited to association with living beings. While it is true that we produce computers and attempt to get AI to run on them, it seems odd to conclude that if we are successful, the AI is intelligent only by virtue of being constructed by us--and should it have somehow come into existence without our intervention, then it would not be intelligent.
I also echo RBH's comments about previous distributed processing work.
Also, when browsing the literature, one should keep in mind that most neuroscientists already assume that biological systems are distributed and have feedback (i.e. are "dynamic"). Experiments often don't reflect this realization not because people think the feedback and lateral interactions and information processing loops don't exist, but because we usually don't have the tools to perform meaningful experiments.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 30. December 2003 15:14
Rex,
As you correctly point out, the definition offered is very broad. This is an inevitable consequence of a definition based on distributed processing. If intelligence is a product or feature of distributed processing, then some of the individual sub-components or sub-processes must of necessity be quite simple. Note that as defined, a simple process is only considered intelligent if it is dependent on and or connected to a very complex LEPS process.
There is currently only one known LEPS system in the universe. The single known LEPS system is the one associated with life forms on earth. If humans create intelligent machines, these machines will be the product of distributed processing perform by humans and would thus be an extension of the ‘life forms on earth LEPS’. It is of course possible that an LEPS system could develop or may already have developed on some planet other than earth.
The definition offered is as you point out very broad. It is not, however, accurate to say that everything associated with life forms is intelligent. Intelligence is defined as characteristic of a some subset of the cause-effect, or input-output relationships associated with life forms. Not all of the input-output relationships associated with a life form would qualify as intelligent.
As have stated on a number of occasions, I do not consider either distributed processing or dynamic distributed processing to be new concepts. DDP has for a very long time been an accepted concept in all applied sciences. DDP is also an accepted concept in modeling of complex processes. Whether or not cognitive psychologists used some variation of this concept is not relevant to the definition proposed or the discussion here.
I also fail to see where RBH’s comments on the professional literature are relevant. The professional literature over the last 50 years, IMO, has failed to produce an explicit, objective verifiable definition which makes it possible and practical to distinguish between intelligent behaviors, processes, and systems and behaviors, processes and systems which are not intelligent. While I am certainly willing to consider any explicit alternative to the definition proposed, I do not see the relevance of recognizing ideas that did not work. I also do not see the relevance of ambiguous ‘might exist in the literature’ definitions of intelligence.
It is worth noting that an explicit, objective, and verifiable definition is an essential prerequisite for any effective scientific analysis in either AI or ID. Unless you start with an explicit verifiable definition, you face the issue of the perpetually moving goal posts. Without an appropriate definition of intelligence, evaluation of a proposed simulation of intelligence or evidence for design by intelligence will be evaluated based on subjective criteria which will always support the conclusion "That is not intelligent’.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 30. December 2003 16:43
warren_bergerson wrote quote:
I also fail to see where RBH's comments on the professional literature are relevant.
Well, in doing real science it's normally the case that one familiarizes oneself with the existing body of knowledge before making categorical claims about it, as for example, warren_bergerson's claim that quote:
But academic scientific and mathematical analysis of complex causal relationships have not kept up with advances in computer technology and systems design. Academic science in large part still operates on the outmoded premise that causation is limited to simple algebraic relationships and stochastic variations of these simple algebraic relationships.
The references I provided, in particular Parallel Distributed Processing, show that claim is simply false. I challenge warren_bergerson to tell us why the approach embodied in PDP is characterized by "...the outmoded premise that causation is limited to simple algebraic relationships and stochastic variations of these simple algebraic relationships. " Don't merely assert it, show it! Let warren_bergerson demonstrate that Rumelhart, et al.'s, approach is as simple-minded as he claims.
Academic scientists are neither as stupid nor as mathematically unsophisticated as warren_bergerson makes them out to be. Some of them are actually pretty smart and they're even fairly knowledgeable about the subject matter they've spent their careers studying!
The various working definitions of intelligence in the literature are each of them incomplete, no doubt. And no definition of which I am aware, warren_bergerson's current effort included, makes that sharp distinction, mostly because there is not a sharp distinction between intelligent behaviors and behaviors which are not intelligent in the world. They're overlapping fuzzy sets, not crisp categories. Hence a definition that makes a crisp distinction must impose an artificial structure on inherently fuzzy phenomena.
I'm not much interested in discussing the various definitions, warren_bergerson's included. Mostly I've posted in this thread to alert non-participating readers to the large literature that warren_bergerson dismisses so cavalierly.
RBH
[ 30. December 2003, 16:55: Message edited by: RBH ]
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 30. December 2003 22:57
Warren, do you mean to say that you define all life (on Earth) and everything produced by it as part of an intelligent system, and that your method therefore does not distinguish between, say, the human brain and the trail of slime left behind by a snail, when it comes to intelligence?
I can't see how I could find such a definition useful. Even if I wanted to define such a thing, I certainly wouldn't call it "intelligence", as the comparison with an intuitive notion of intelligence is more confusing than illuminating.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 31. December 2003 09:28
Rex,
It is generally agreed that intelligence is some type of complex information processing. In developing computer programs involving complex information processing, one of the basic tricks used is to break the complex processing into dispersed processing modules and ultimately into simple process and lines of code. The evidence would appear to support the assertion that intelligence in biological systems involves this same concept of dispersed processing modules. The definition proposed here simply expands this concept to suggest that all the intelligent processing associated with intelligence involves a single dispersed processing system. This concept is useful because it requires that models of intelligent processes recognize different forms of communications between modules as well as the processing algorithms in the modules.
The concept of communicating dispersed processing modules has proved useful and practical in developing complex computer programs. There are good reasons to expect the concept will prove useful in the analysis of intelligence.
As I stated yesterday, intelligence is being defined here in terms of complex causal processes or information processing not in terms of physical objects. The proposed definition might not attribute intelligence to a slime trail, but it would attribute intelligence to some of the information processing associated with snails and cells of snails.
One of your comments yesterday suggested it would be inappropriate to attribute intelligence to the complex processes involved with generating energy from sugar. The definition being proposed recognizes that an act of intelligent behavior, like learning algebra, is dependent on very complex preliminary processes. Having an energy source available to perform the processing associated with learning algebra is an essential prerequisite for learning algebra. Developing an energy source and developing the capacity to control the energy source involves very complex information processing which would be part of a LEPS.
RBH,
There have always been debates about whether applied science leads theoretical science or theoretical science leads applied science. The fact that on a specific issue or set of issues there are inconsistencies between the two types of science does not suggest that one group is stupid. In the area of AI today there appears to be a significant gap between what applied science programs can model and simulate and what theoretical scientists are willing to recognize as possible. The definition of intelligence offered here is simply based on concepts system design experts use in practice to model and simulate intelligent processes.
As you concede, academic scientists are not succeeded in produces a definition capable of differentiating intelligent and non-intelligent behaviors, processes, and systems. I suggest one reason academic scientists have failed to produce an appropriate definition of intelligence is their failure to recognize complex causation in the form of DDP and goal-directed DDP.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 31. December 2003 20:03
warren_bergerson wrote quote:
The definition of intelligence offered here is simply based on concepts system design experts use in practice to model and simulate intelligent processes.
References? Please direct us to just one reference that is an example of that use. What "system design experts" are you talking about and where are specific examples of that use?
warren_bergerson further wrote quote:
As you concede, academic scientists are not succeeded in produces a definition capable of differentiating intelligent and non-intelligent behaviors, processes, and systems. I suggest one reason academic scientists have failed to produce an appropriate definition of intelligence is their failure to recognize complex causation in the form of DDP and goal-directed DDP.
As the PDP reference I supplied above demonstrates, warren_bergerson's claim about academic scientists failing to recognize complex causation and distributed processing issues is false. Period. Those two volumes alone contain dozens of papers reporting actual research (as opposed to armchair speculations) on distributed systems displaying the properties of intelligent systems.
Nor have AI people, applied researchers, or systems design people provided the kind of definition warren_bergerson thinks necessary. Nor has warren_bergerson provided such a definition, as Rex Kerr's remarks demonstrate. I challenge warren_bergerson to supply one actual reference in the professional literature of any appropriate discipline that supports his claims.
RBH
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 02. January 2004 09:40
One of the useful features of the proposed definition is that it makes it possible to use mathematics and mathematical analysis to address a number of interesting scientific and philosophical issues. Dembski’s Irreducible complexity and Neo-Darwin RM&NS provide to useful examples.
Dembski’s IC concept asserts that it is prohibitively improbable to produce certain types of complexity using a simple variation and selection process. The proposed definition, however, suggests that the existence of goal-directed DDP pre-processing should make it possible to produce very complex phenomena using very simple final stage variation-selection processing. It should be possible to develop a computer program/simulation which consistently and reliably evolves ‘IC’ features.
On the other hand, it should be possible to demonstrate mathematically that simple Darwinian variation-selection processes in the absence of complex goal directed DDP pre-processing would have little or no ability to produce evolutionary change.
The proposed definition of intelligence, it is suggested, makes it possible to use objective, verifiable mathematical analysis to evaluate some of the controversial concepts and hypotheses involved in the ID vs Darwin conflict. Note that I am suggesting is should be possible to develop models and simulations to test these concepts and hypotheses, not that such models have been developed. In summary, one of the major benefits of the proposed definition is to make it possible to address issues using objective, verifiable criteria and thus replace the subjective, authoritarian, conformity with the literature criteria.
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 02. January 2004 12:50
Warren: On the other hand, it should be possible to demonstrate mathematically that simple Darwinian variation-selection processes in the absence of complex goal directed DDP pre-processing would have little or no ability to produce evolutionary change.
I am not sure how one is going to achieve based on using intelligence as the relevant measure? In addition to evolutionary processes being teleological in nature (selection).
Mark Toussaint, one of the authors of the infamous NFL theorems, published his thesis
M. Toussaint (2003): The evolution of genetic representations and modular neural adaptation. Submitted (in April) version of my PhD thesis
quote:
Facing the challenge to understand intelligent biological systems, a first step is typically descriptive, i.e., aiming at a precise description of a system’s functionalities and mechanisms. This is true for many branches of biology when describing the subtle mechanisms of natural life, for many branches of Neuroscience describing in detail neural mechanisms, and eventually also for classical approaches to Artificial Intelligence which try to capture intelligence on a formal descriptive basis.
However, a purely descriptive approach might neglect that all the intelligent systems we find in nature are the outcome of adaptation processes, namely evolution and neural adaptation. This puts a crucial constraint on what systems can possibly exist in nature, in particular how they must be organized and structured to allow for adaptation. This calls for a theory on these adaptation processes themselves as the basis to understand natural intelligent systems. Why does the need for adaptation induce a structural constraint? The point is that nature does not invent intelligent systems directly. All of a system’s functionality is represented indirectly by genes or neurons and adaptation takes place on the level of these system parameters. The way these parameters encode the system’s final features is highly non-trivial. For example, engineers often profit from precise descriptions of nature and have adopted many techniques. However, when they design systems they usually directly describe and design the system’s final functionality; they hardly adopt nature’s strategy to use an indirect representation of functionalities by genes or some
equivalent. This leads to a problem that becomes particularly apparent when the goal is the design
of an Artificial Intelligence. Early approaches to Artificial Intelligence were largely descriptive, i.e., tried to first directly describe what intelligent behavior is in terms of behavioral rules. These rules could then be implemented on a computer. [bBut actually, a precise direct description of intelligence in terms of rules is very difficult. One realizes that whenever one formulates a behavioral rule, one needs to presume some vocabulary of situations and behaviors to formulate the rule [/b](e.g., “when in situation A execute behavior b” presumes that it is well-defined what situation A and behavior b are). Nature developed a elaborate representation to encode intelligent behavior, namely highly structured neural systems. This representation is the outcome of a long adaptive process. The crucial question is how the adaptation processes succeeded in developing these indirect representations that are so beneficial for functional adaptation.
[ 02. January 2004, 13:04: Message edited by: Pim van Meurs ]
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 02. January 2004 14:23
Warren, your mathematical definition is broken.
Something is intelligent if:
quote:
A behavior, process or system is characterized as intelligent if and only if the behavior, process, or system is dependent on, associated with, and a product of a complex and extended preliminary processing called ‘learning efficient problem solving’ (LEPS).
And something is mathematically LEPS if:
quote:
The mathematical definition of LEPS is expressed as a computer program expressed in pseudo code. This LEPS program is a program capable transforming a set of ‘naturally occurring stimulus-response, input-output, or cause and effect relationships’ into a set of ‘stimulus-response, input-output or cause and effect relationships’ associated with an intelligent behavior, process or system.
(Emphasis mine.)
Well, fine, LEPS is associated with intelligence and intelligence with LEPS, but this isn't going to help us much when trying to determine what is intelligence and what isn't! Perhaps you are trying to use the LEPS definition to generate a closure of the set of things that are considered intelligent?
As it stands, the definition is circular, and it's such a small circle as to not really be instructive.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 02. January 2004 14:48
warren_bergerson wrote quote:
In summary, one of the major benefits of the proposed definition is to make it possible to address issues using objective, verifiable criteria and thus replace the subjective, authoritarian, conformity with the literature criteria.
I take it then that warren_bergerson cannot supply the references I requested to support his claims about what "expert AI systems designers" do. Instead, we are to rely on warren_bergerson's authoritarian pronouncements. I see no reason to do so, and so will withdraw from this thread.
RBH
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 03. January 2004 08:52
Rex,
The proposed definition most certainly is not circular. Furthermore, the definition provides a clear, objective and verifiable method of determining if a defined behavior, process, or system is or is not intelligent as follows:
Given a defined behavior or process X (defined here means can be expressed or modeled by an information processing algorithm) then X is intelligent if all three of the following criteria can be satisfied:
1. X is ‘connected to’ or dependent on living systems.
2. X does not occur in nature except connected to a living system. and
3. The algorithm expressing X can be generated from non-intelligent processing by a mathematical algorithm belonging to the class LEPS.
You may need additional information as to the meaning of ‘connected to’ and ‘LEPS’ to actually perform the evaluation, but that information does not depend on whether X has been defined as intelligent.
PIM- I don’t see what point you are trying to make.
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 03. January 2004 12:24
It simple, rather than approaching the issue through a vaguely defined concept like intelligence, a much more fruitful approach has been chosen by Toussaint. Since biological procesess can be teleological, or goal directed, I do not think that your approach will be useful in distinguishing between truely intelligently designed and natural designed.
Toussaint's approach seems well founded in mathematics and helps understand intelligent biological systems.
Toussaint argues actually, a precise direct description of intelligence in terms of rules is very difficult. One realizes that whenever one formulates a behavioral rule, one needs to presume some vocabulary of situations and behaviors to formulate the rule
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 03. January 2004 14:07
I suppose Warren's mixing his conclusions in with his definitions then--that's the only other explanation for the use of intelligence in the supposed definintion of LEPS.
I still haven't the foggiest idea how I would rigorously use these definitions.
In any case, I believe that I'm done with this conversation. The posturing-to-content ratio is simply too high. Or perhaps intelligence-(implemented)-as-distributed-processing is a new idea to Warren, but it is not new to me.
In any case, an interesting idea which is underdeveloped is not useful. Promising, perhaps, but not useful. I'll wait until I see some more solid glimmerings of genuine utility.
IP: Logged
|
|
|
Printer-friendly view of this topic