Question for Artificial neural network (ANN).....

Wendy

Joined Mar 24, 2008
23,415
I have to disagree, first, I dont really think you can call the brain a computational system, but secondly even if you call it that, no turing machine can ever simulate an analog signal or function, because a turing machine is a discrete engine and can never calculate with continuous numbers. A touring machine could aproximate a neuron, but never truly emulate it; it simply lives in a completely different world.
I've had this argument with others, I think we'll have to agree to disagree for now. Since it is named after a individual (who was treated rather shamefully by his government) it is with a capital T, as in Turing. He was gay when it was a crime to be so, and was chemically castrated by the courts. When he started growing breasts he committed suicide.

This after his ground breaking work during WWII, several sources claim he ended the war early and saved many lives doing so. The British government has formally apologized.

Fact is, we are not even close though. To me this says we are missing something important.
 

Wendy

Joined Mar 24, 2008
23,415
Like I said, there is no good theory for true AI. Personally I feel if it is ever invented it should be call synthetic intelligence, since it will be the real thing.

Want a Nobel Prize? Figure out the theory. The reason no one is replying is we do not have the answers.
 

bwack

Joined Nov 15, 2011
113
Why we use function like sigmoid or other function??
Sorry for the late reply.

It's not the sigmoid, but a sigmoid function. A sigmoid function has the S shape and maintains the output values between two limits. If you're dealing with probability estimation with a artifical neural network (ANN), then you want the output to be contained between 0 and 1. One of the hyperbolic functions (tanh) can do that, and there is another sigmoid function that the output can be between -1 and 1. Another reason for this is to meet requirements for the algorithms I explain below. To use them, it is required that the function to be differentiable and convergent. The area under the graph when the input goes from minus infinity to positive infinity is equal 1.

In the mid 80s when the backpropagation algorithm made a break-through for feed forward multi-layered ANNs, the gradient descent algorithm was used prior to that era (50-80s). These algos are similar; backpropagation is a generalization of the gradient descent. The problem with gradient descent is that computational power required goes through the roof quickly when the size of the network increase.
For learning n neurons you need:
1. a power of n computations for gradient descent.
2. a multiple of n computations for backpropagation.
A multiple (np) is much easier to appreciate than the power of n (p^n) where p is one computational unit of learning. Now it was possible to have larger ANNs. Even with the computers of today this is a great advantage. (the BP was actually reinvented, because someone had allready found the algorithm back in 1969, but ANNs was ignored until the mid 80s. You have to realize how slow computers were back in the late 60s. )

I dug a little deaper last night for uses, and I followed the work of one of the guys studying the use of ANN for use with cognitive signal processing in the field of radar. Dr. Haykin
http://soma.mcmaster.ca/haykin.php
Have you ever wondered why radar screens are so damn clean ? How do they filter the oh so much clutter that returns from the radar antenna? At sea, how do they know it is an obstacle they are approaching and not just some scattered ice ? From reading on a paper last night from the 90s (ill dig it up) in addition to the many filters used to generate the radar image, an ANN was attached at the very beginning of the filter process to classify patterns to mask in or out of the image. Think about it, most human made objects are not very random in texture, they have large smooth surfaces where as in nature everything seems very random. It turned out that ANNs can be learned to see these human-made patterns. It was also used to clasify ice too, and also for use with satelite radars to measure the size of sea ice on earth. Just food for thought :)
http://www.seaice.de/

I generally don't like to participate in discussions of wether ANNs are "intelligent" or not. For me they are mathematical models that can be thought to see patterns, however powerful they are, they are very limited in use in the real world. The limited use part may be the reason why this field is not so popular and you get not so much answers.
 
Last edited:
Top