|
Author
|
Topic: Suggestions for an AI Project
|
Douglas
Member
Member # 641
|
posted 09. February 2003 11:44
gedanken,
"Recognition" is not the same thing as Creation, or even as understanding. I wasn't proposing that any neural network could reason about what constituted "design", or why it recognized "design".
Of course, if a neural network system could recognize "design", at least within a limited "environment", then what would prevent it from "concocting" an instance of "design" which it would recognize as "designed"? I don't know, offhand. But in the sense that the neural network system was merely "mimicking" instances of "design" which had their "origin" with humans, whatever "design" it would produce would be "human mimickry", rather than "natural", I would say.
IP: Logged
|
|
kyle7
Member
Member # 191
|
posted 17. February 2003 21:59
Gedanken,
I think someone can program an AI system to detect design. Of course you need intelligence to do the programming.
Frances asks the following:
quote:
Do you have any suggestions how to deal with the concept of specification in the AI algorithm? Specification seems to be a very important aspect of ID inference and yet it also seems to be the most subjective one. Do you propose any ways of inferring an objective specification detector?
The issue of specification is rather simple at one level. You would use statistics coupled with human intelligence. For example, you could take any sensor and take samples from a "natural" system and build up a data base of information that describes what is natural. If your system is a visual system, say for a missile detection system, then you could sample pictures from the sky to build up a data base that describes what is natural. Multiple sampling could be done from a single picture. You could sample for shape, color, or other criteria (e.g. change of shape over time, for video). If your sensors were pressure sensors on the fingertips of a robotic arm, you may sample by taking pressure readings as the robotic hand scrapes through a sand pile. You could "teach" the robot to find an object in the sand and discern if it was a natural rock as compared to a round marble, watch or bottle cap.
You could also use physics coupled with statistics in developing your data base, but this would require more effort. For example, you could use a fluids code to study sand formations. Based on this information, you could develop your functions that describe what is deterministicly natural. One could use this function for evaluating if a sand formation was natural or intelligently designed.
Rex Kerr does offer some reasonable advice. Projects can easily become a monster hard to slay if you take on too much! But, I do think an exceptional student could think up an idea and use Dembski's approach with success. The student would have to be wise in using his/her available resources.
One final note. You would have to play around with the cut-off number that describes if the object is designed. By testing the model you could determine the best cut-off number. Perhaps, you could setup the system to evaluate a probability of it being designed (e.g. 95% probability of ID). You would look at all "objects" with greater than say 80% design for closer analysis. Any sensor readings outside of the "probability space" or on the edge would be objects potentially suggesting design. I say
"potentially" because with a simple project like
this you don't have a lot of time for sampling.
Here, I am fusing the random and the deterministic together by using statistics, which some may think is a bastardization of Dembski's design filter.
Micah, wishing you well on your project.
Kyle
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 18. February 2003 00:56
Kyle7, Douglas,
Of course one can do a simple pattern match for existing known (or described) “designed” objects. Military’s “smart weapons” in research already do this, for example, providing an example that it can be done. (The problem with this is it is not appropriate for Micah’s AI class, as it is a very narrow sub-domain of AI called “pattern recognition” and not relevant to the class project, however interesting it might be.)
However if you consider the more generalized case of aspects of shapes for the search to look at, then we have a more interesting case. But one of reasonable size would have problems with things like snowflakes, with rock crystals, with round river rocks, with “stone circles”, as compared to manufactured items with similar characteristics in or of their parts. Teasing this out becomes an extremely difficult project. (And this applies to Douglas’s point as well as Kyle7’s)
It is in fact getting people to recognize the degree to which “common sense” is not easy or “common” is part of my point. If you consider this to be a reasonable challenge for the AI system, then you agree that what we would consider intelligent behavior can be done with reasonably simple computational systems. (And I find this to be a contradiction to other precepts of ID itself -- what would your position be on physical systems doing “design” in nature, for example?) And if you don’t consider this a reasonable challenge, then we are contradicting the stated case by both Kyle and Douglas.
My main interest is in getting people to rationalize their views of what a physical system (like a computer) can do, and their views of the complexity of recognizing any significantly interesting cases of ID beyond simple pattern matching to existing pictures already declared “designed”.
IP: Logged
|
|
Douglas
Member
Member # 641
|
posted 20. February 2003 13:20
gedanken,
I think you misunderstood me. When I mentioned "teaching" a neural network system to recognize "design", I was thinking along the lines of teaching it to generalize the characteristics of "design", at least within a limited domain (say, of visual shapes/objects). Not exactly the same thing, I don't think, as "pattern recognition", though close. Perhaps it would need some "nested" neural networks. Perhaps I'm revealing my ignorance.
IP: Logged
|
|
|
Printer-friendly view of this topic