New developments in AI

Discussion in 'Intelligence & Machines' started by Rick, Oct 18, 2006.

Thread Status:
Not open for further replies.
  1. weed_eater_guy It ain't broke, don't fix it! Registered Senior Member

    Messages:
    1,516
    I'm kinda dumb with regards to this field.

    Why isn't it as simple as setting up a number of programs to function like neurons, and have those programs form a network similar to a neural network?

    Is there a good link or something that could explain the difficulties of AI in that respect? Is it the programing of said "neuron"? Is it the interconnection between them?
     
  2. Google AdSense Guest Advertisement



    to hide all adverts.
  3. mackmack Registered Senior Member

    Messages:
    123
    a neural network isn't a very good model of a brain. it acts as a universal ai that can be applied to things like cars, face recognition and business statistics. but the network itself is flawed because information is "hidden".


    the information that denotes its intelligence can't be seen by the user therefore its useless. in the neural network there is a part called hidden layer. they don't call this hidden for nothing.

    on the other hand i define the hidden layer in my program. this will allow greater control over what gets inputed/outputed and modified in the neurons.

    this will also allow the network to change only specific data in the neurons and not the whole network.

    its a complicated peice of software but it can be understood in a simple way. here is the link.

    http://www.humanlevelartificialintelligence.com

    there you will find the information about how i implement a network that does exactly what a neural ntwork does but the coding is done completely different. to understand how it works all the material in the website is to be understood.
     
  4. Google AdSense Guest Advertisement



    to hide all adverts.
  5. weed_eater_guy It ain't broke, don't fix it! Registered Senior Member

    Messages:
    1,516
    very cool, you got a youtube or something with this software in use? i wouldn't know how to run something like that by my own
     
  6. Google AdSense Guest Advertisement



    to hide all adverts.
  7. zenbabelfish autonomous hyperreal sophist Registered Senior Member

    Messages:
    961
    'the network itself is flawed because information is "hidden".'

    Hi Mack...this is an interesting comment from a neurobiological perspective. Why is the information 'hidden'?
     
  8. mackmack Registered Senior Member

    Messages:
    123
    why is the information hidden?--zen

    they have three layers in a neural network:
    1. input layer
    2. hidden layer
    3. output layer

    the input layer is the data that goes into the network. for example if your AI program is a face recognition then the input data is a picture of someones face. if your ai program is a fingerprint program then the input data is a picture of a fingerprint.

    the output layer is the classification of the data from the input layer. if you wanted an ai program to recognize numbers then the output layer is the numbers. an image of a number will be the input data and the classification of that image is the output data.

    example--a handwritten image of a number 4 is the input data. the ai will classify this image as the number 4.

    the hidden layer is the stuff that denotes the strength of "all" the data in the network. it uses math to calculate how strong one input data is with each of the output data (it is hidden because its pure math). This is why the neural network has to train all the data in the network in order for it to work properly. the human brain works the complete opposite, it only trains specific data in the network and not the entire network.

    The most popular ai tool in todays market is the neural network. The universal AI program is something that will solve all the problems of the neural network. it trains only specific data in memory. the scaling such as storage is optimized. actually, you can store infinite data in the network because i use a form of fuzzy logic. the current networks can't store more than 10,000 different face images. mine can store infinite face images. the retrieving of data is faster than anything in the market. etc.
     
  9. zenbabelfish autonomous hyperreal sophist Registered Senior Member

    Messages:
    961
    Wow - that is amazing...thanks for taking the time to explain so succinctly. I need to do some research here.
    My interest lies in whether the human mind has a cognitive 'blindspot' analogous to the blindspot in visual perception. I wondered if there was a parallel or analogy to the 'hidden layer' in neural networks.

    As you describe, the hidden layer is hidden because it is pure math and calculation...does this also mean that it is hidden as it is intrinsically part of the process of the operation of the neural network?

    I guess a metaphorical question would be: Is it the movement of the river or the riverbanks?

    Apologies...bit of a wild card...thanks.
     
  10. zenbabelfish autonomous hyperreal sophist Registered Senior Member

    Messages:
    961
    Sorry better metaphor...is the hidden layer like 'an eye that can see everything but itself'?
     
  11. Lomion Registered Member

    Messages:
    18
    I think I can answer the "blindspot" question.

    Yes. The human mind does in fact have a series of blindspots that prevent it from being as universal as possible. These are called "sets" (AKA someone is "set" in their ways) in psychology. It is merely a filtering of possible options that a person tends to ignore. Sets reduce creativity in some cases, but enhance efficiency in others.

    However, a set to an AI would probably not be a good thing. But then again, There's probably sometime when it would be appropriate. Also, I could be blatently wrong. (As noted earlier in the thread. Thanks for correcting me dinosaur!)
     
  12. zenbabelfish autonomous hyperreal sophist Registered Senior Member

    Messages:
    961
    Thanks Lomion...very interesting. Sets sound a lot like the filters described in buddhist psychology too. This a real big help...sorry for going OT.

    Please Register or Log in to view the hidden image!

     
  13. kmguru Staff Member

    Messages:
    11,757
    Wait a little bit...I mean a few years...for dual core to go to quad core (already there) to perhaps go 1024 cores with a terabyte DIMM - then you can teach the computer to do something...

    Microsoft has an AI research lab, but the boys are asleep...too many pizzas...even their Bayesian network has a long way to go...

    I wonder if quantum computers can help...what is the latest?
     
  14. mackmack Registered Senior Member

    Messages:
    123
    when i was in college it took me 1 week to actually understand what the neural network does. After i understood the concept, i told myself how ridiculous the creator of the NN was. The reason is because the concept behind it is so simple but the creator made the concept hard to understand because of all the math involved. even explaining how to apply the neural netowrk to something like face recognition was complex.

    so basically, i just took the math out of the neural network and worked on my own model. what i found was that i can define the hidden layer, not using math, but using art. when i defined the hidden layer i also found how to universalize the input layer and also how to universalize the output layer. thus thats how my program got its name: universal artificial intelligence program.

    as far as the hidden layer is concerned the computer dont' even know what the inner workings are, like i said its pure math. the math involved is just to calculate the strength between the input layer and its relationship to each of the output layer. eg. if someone writes the number 7 in a funny way, the strength of that input will be strongest on the output number 7. but because the number 7 was written in a funny way the computer also strengthened the connection to number 9. 7=80% probability 9=40% percent probability.

    only one of the categories in the output layer can be activated. since 7 has the highest probability then that is the output data that is activated.
     
  15. Lomion Registered Member

    Messages:
    18
    kmguru: Forget quantum computers. They will never ever be able to make one of those things powerful enough to do the calculations. The best way would to be to go with a linux cluster.

    For those that might not know about clusters, they work by making a network of computers act as a single system. Usually, one (or very few) computers are given a kind of "head" computer. That is, they are the ones that coordinate the whole network. They take a program or math calculation and send a piece of it to each computer. Every other computer then calculates the result of that piece, and returns the value. In theory, you could have up to 1024 individual cores, with terrabytes of DIMM. These systems are easier to maintain, and cheaper to build than one of the big-rigs.

    They still cost quite a bit to make that big, and I don't think networking standards are up to that daunting of a task yet. Plus, you have to have the people who would want to use that system agree on how to use it.

    mackmack: Yeah, I've been doing some research on neural networks. I don't like the standard either. I'll probably try to understand it, and when I'm finally sick of it I'll quit and decide to write my own version of it. Good source code on this stuff is hard to come by!
     
  16. Dinosaur Rational Skeptic Valued Senior Member

    Messages:
    4,592
    Lomion: I think you are right on the money. It is my best guess that the potential capabilities of quantum computers have been wildly exaggerated. At best, they & DNA-based computers might be usable for very specialized problems.

    The clusters you describe are very usable for a small class of problems which can be broken up into little pieces. The current architecture runs the same program on each CPU in the cluster. This is great for solving certain partial differential equations, cracking keyword codes, and solving some other very specialized problems. The architecture does not seem suitable for implementing true AI.

    As I previously posted, current computer architecture will not implement true AI, even if CPU speed, memory, and disk space are increased by orders of magnitude.

    Until we have a better understanding of how the human brain functions, we are unlikely to be able to design or program true AI.
    • The human brain does not seem to have separate hardware & software functions. Much of what we naively would view as software seems to be implemented by rewiring what one would consider hardware in the brain.

    • The human brain does not seem to have anything analogous or similar to an indexed database.

    • Memory seems to be implemented by rewiring what one would consider to be hardware, rather than storing chunks of data in reusable memory allocated for data storage. In the human brain, there does not seem to be a clear cut distinction between data memory, memory containing software functions, and what we would view as hardware.

    • Our brains seem to function via the use of a vast network of circuitry analogous to a large number of hyperlinks operating in some incredibly parallel manner. For example, when you perceive a rosebush, the data is scattered over many parts of your brain. The data about shape of the leaves, flowers, and supporting trunk & limbs are stored in area of the brain (probably not all together in a specific part of that area); The color data is stored elsewhere; the verbal data for color & shape is in another area; The data required to drive motor nerves for lips & tongue to allow talking about a rosebush are in still another place; Data about related fauna & history of a rosebush are stored in another place; Data about bees/birds which pollinate a rosebush are stored in still another area (linked to all sorts of data about bees & birds). When visual data enters your eyes you collect all this data in order to create your perceptual image and prepare to think & converse about rosebushes.

    • Consider your perceptual image of a group of brick buildings. Consider how your perceptual image correctly responds to your walking or running as you look at the buildings. Those images seem to be built in great detail in real time. When I look at a brick wall, I seem to have data allowing me to count the individual bricks, assuming that I could make an exact copy of what I think I see. I do not think that the retina, optic nerves, & brain processing functions are actually capable of collecting and processing the necessary data in real time. In particular, I wonder about the band width of the optic nerves. Our brain creates the perceptual images of a brick wall (or a large tree with thousands of leaves) using some visual input data (less than actually required for the image), memory of previously created images, and a lot of damned complex data processing done in parallel.
    Our current computer architectures and AI software just do not seem capable of doing the job done by our brain. Furthermore, I do not think that more memory, more disk space, and faster CPU’s will do the job without some major break through in software & hardware design.
     
  17. URI IMU Registered Senior Member

    Messages:
    729
    As an aside here, I am wondering who would be up to a team AI programming project ?

    The programming language should be very low down, something like C. Programming ability is desirable but not necessary.

    If interested you could register at http://omegafour.com/forum/ and post your interest in the last forum of the opening page entitled "Requests"

    I will then give you privileges and allow you into a secure AI forum.
    Thanks.
     
  18. URI IMU Registered Senior Member

    Messages:
    729
    >> I do not think that more memory, more disk space, and faster CPU’s will do the job without some major break through...........

    The major breakthrough is mathematics. The type of math necessary is unique and compact. All the information of the world could be encoded into one mathematically analysable number.... similar to the massive storage ability of the DNA code. After all mental power is chemically/electrically driven and there really is no mystery, since mathematics is the only game in town.

    The program to empower thought is very small but mathematically very powerful, it is the data capacity that allows intelligence.
     
  19. Lomion Registered Member

    Messages:
    18
    URI: The math we have at our disposal isn't quite up to par, true. However, it's the only tool we have at the moment. Ancient astronomers didn't have the means to physically measure the distance of things like the Earth's circumference, or the distance to the moon. But with what they had, they were able to calculate how far it might be. Then modern physics came in, and got us what we needed better. We have to be the ones to start it, or it won't happen.

    I can understand what you're saying, but you haven't offered any kind of solution to the problem. Where would this math come from? Can you cite any work being done to fix the problem? What will this new math be like? Personally, I don't need too many details. A vague idea would be nice, though.

    Dinosaur: You're right about the clusters in most respects. However, if you were to implement a neural network on them, theoretically, wouldn't it be more effective to actually have a network? It wouldn't just be a software design anymore, but a combination of software and hardware that spreads out the system a little more thinly.

    However, I still have to agree with you on the software/hardware point you made. Even with the cluster as such, it's still not the same as a human brain. Nothing short of a biological computer would be able to compare. (I don't wanna bust open that can of worms. I'm just using it as a reference to very advanced technology of which I don't think I'll see in my lifetime. Please don't diverge into it unless neccesary!

    Please Register or Log in to view the hidden image!

    )

    Also, I've noticed a lot of comments about the optical nerves, and data compression of that area. Again, I agree that we won't be able to properly understand how to make a good AI program without a better understanding of ourselves. The best thing to do, would be to have a better computer that can handle all the extra data that we can't compress digitally. I'm open to new ideas or alternatives, though.
     
  20. URI IMU Registered Senior Member

    Messages:
    729
    >> you haven't offered any kind of solution to the problem

    I have solved the math... I call it symbolic math
    and NO, I will not specifically talk about it, nor release the algorithms.
    AFAIK, such math is not known on this planet.

    As an example, take the word "cat"
    a computer may match c+a+t = cat.... so what hash function do you use for "cat", whatever it is it will need a look up table.... and so forth.... so messy and non specific

    Symbolic math has no look up, what cat is as a word is encoded, as well as what cat is as a thing, into a unique (specific) analysiable number, the representation is as if it was the actual input ... and all else you determine about cat can be progressively added, with no loss of coherence.

    That is as far as I will go. basically when you bring up a concept all associated concepts are there in the single numerical representation.... and of course logic is in the deconstruction/reconstruction of this always unique number. IMO this math is the way the brain functions.
     
  21. Lomion Registered Member

    Messages:
    18
    Breaking news! An article has just recently been published stating that a silicon replica of the cerebral cortex has been made. It's going to be an array of 256x256 transistors that will run 300 Teraflops. That means real time interpretation! There's more info on it at:

    www <dot>
    technologyreview <dot>
    com/
    read_article <dot>
    aspx <?>
    id=18164&ch=infotech

    Apparently, I still can't post links yet...
     
  22. mackmack Registered Senior Member

    Messages:
    123
    "You're right about the clusters in most respects. However, if you were to implement a neural network on them, theoretically, wouldn't it be more effective to actually have a network? It wouldn't just be a software"--dinasaur

    the software that simulates a brain has to be in a network. well, at least in my program it serves as a semi-network.

    the information in the brain is global, that means all the data is connected to each other one way or another. thats how data association works.

    the problem with the neural network is that it has a fixed set of neurons in the brain. if one extra neuron is added that entire data in the network has to be changed based on their wieghts to each other. this is like using brute force to find a match in a fixed number of items. using bubble sort is brute force (like the neural network). but its much faster if you use quicksort.

    the nueral network is done in a linear way while the brain uses multiple ways of doing things. thats why the neural network isn't a very good tool to built a human brain.

    i solved this problem by building a computer/machine/software that can store infinite amounts of information and to connect all the information in a network fashion. yes, you heard me, a computer that can store infinite amounts of information. and the searching algorithmn is quick too so that regardless of the number of data in the computer the computer will get that information quickly.

    the secret is the fuzzy logic that i have included which is created by the 4 deviation function i mentioned on my site.
     
  23. Dinosaur Rational Skeptic Valued Senior Member

    Messages:
    4,592
    Some interesting claims being made by posts to this thread.

    All the information about a concept like cat encoded in a single number, reminding me of Goedal numbers. An infinite amount of information stored in some system.

    I await the articles to be published in some prestigious journal. I wonder if there is a Nobel or other prize for such discoveries.
     
Thread Status:
Not open for further replies.

Share This Page