Hi all,

I been resarching the internet seeking information about neural networks, few years ago i decided to start a project wich that solves problems that lineal computer systems cannot easily solve, i discovered Artificial Neural Networks i have implemented some models and algorithms, however i would like to know if there are more news about this amaizing technology, all models and algorithms have more than 10 years. There more? and where may i found they?

In my understanding, for realy realistic neural network like in human brain to scientist is very far and computer power is not enough, becouse neurons working in diferent way and maybe neurons are himself powerfull computer. In my opinion scientist doing mistake, that they want solve neurons equations with diferencials and integrels. I think, that need just imagine, that neuron is analogue computer with some finity precision with say 100-1000 digital levels of signal.
If you interesting how much speed have human brain and how much power have computer, then need simple calculations. Computer have about 10^9 transistors and 10^9 Hz. So all one computer power is 10^18 operations. Human brain has about 10^11 neurons and about 10^15 sinapses. Each sinapse have about 100 resistivity levels. So 10^15 * 10^2 = 10^17 operations. So like you see human brain power is almost equal to computer power. But each neuron working more dificult than one transistor and that becouse human brain power is perhaps much bigger than computer power.

I little bit correct my previuos statement.
So processor power is 10^18 ops.
Human brain power is 10^17 * 10^11=10^28 ops. So to calculate human brain power need multiplicate number of neurons times number of sinapses times number of sinapses strenght levels (sinapses resistivity becoming smaller if more frenquently electricity going through them).
So human brain calculation power is 10 bilions times bigger than say intel core2 quad.

There seems to be two different disciplines associated with neural networks.

One is the digital approach, such a hopfield net, and the other is the analog approach, such a carver meade's subthreshold cmos vlsi spproach.

One providea a general tool for solving real world problems (hopfield) and the other provides a tool for understanding cognitive processes (and to a lesser degree solving general real world problems, although the potential is there)

I think (not an expert) the hopfield approach (least mean squares convergence) is generally a pattern recognition approach, where training a system with learning pattterns can help recognize or analyze new patterns for similarities. One of my favorite real world examples was a hard disk drive system where a research neural network learned the optimal tradeoffs in detecting high bandwidth data from large number of disk heads given all the head, media, and amplifier production variables (very difficult problem with many many variables). The single learned response was then used in all of the production drives, resulting in a substantial speedup.

The Carver Meade ("Analog VLSI and Neural Systems")approach is (well, used to be, anyway) trying to mimic human processes (vision, hearing, etc) with human like models using sub-threshold cmos devices (which have a neural like logarithmic response). Also, studying the simplistic neural processing systems of spiders or other animal gives insight to how dramatically neural processing affects behavior. (Nerve Cells and Animal Behavior, David Young).

So, there are two distinct sciences here, I think. I find the latter more interesting personally because of its philosophical insights.

Trying to find and apply metrics to benchmark the human brain is really quite meaningless, especially if you begin to consider quantum effects as part of the process. Also, the human process has a huge arsenal of built in specialized co-processors. For example, interpretation of the acceleration of falling objects is build into the neural processing of data from the eye before even reaching the brain. The built in auditory processing (masking effects, etc which are exploited in mp3 and similar compression algorithms) is another example. Just performing those two processes in real time would bring generalized computer systems to their knees.

Maybe Neural Networks are just efficient with recognition and in some cases solving problemas, maybe the real solution after all is genetic programing.

The fastest chips I have read about run at around 724 gHz. I think eventually grids of virtual quantum optical supercomputing hybrids will speed things up by orders of magnitude beyond what today would be considered insanely and rediculously impossible. I suspect a breakthrough is coming down the pipe so to speak, where all of the existing technolgy we have today will be considered funny stuff. Using visible light would allow us say 540 teraherz speeds. Gamma-ray containment fields would of course provide the fastest speeds. I do at times wonder if the LHC will be able to perform computations.

PJParent001, you a little bit with your fanatsies. Chips will never work on about 700 GHz, becouse need that such chip would be with very small size (and hence small number of transistors) and one tranzistor must be 10000 nm size or in over case such transistor will burn and will be broken or will not work becouse all signal became cpacity potencial, becouse higher frenquency signals are more sensitive to small capacitors effects, becouse every object is like very small capacitor and then all energy woul become capacity of 700 GHz transistor, so 700 GHz may be not can be very small. In future processor after say 1000 years would be maximum 1000-1000000 times faster, but it's still not very enough for brain simulation.
I think need begin real neural networks simulation from worms size neurals networks, which have only about 200 neurons. Need creat small legs, receptors and all things which have such worms and then try to creat artificial worm with 200 neurons. Nobody already create such worm, but wanna simulate human brain

Perhaps.

Engineers set new world record in generation of high-frequency submillimeter waves
http://www.physorg.com/news95937386.html

>Nobody already create such worm, but wanna simulate human brain

We already have humans, computers, and robots.

What if i use a parallel system? A Neural Network works parallel, im thinking to use opamps as a neuron ,resitors and diodes as weights, that may work, althought there is the problem of evolution, the system wont be able to envolve.

Chilly chip shatters speed record
"The prototype operates at speeds up to 500 gigahertz..."
http://news.bbc.co.uk/2/hi/technology/5099584.stm

To train it, you have to have adaptive feedback that changes the values of the resistors.

http://en.wikipedia.org/wiki/Neural_network_software might point you to somthing that will help with your comprehension. Also, a more fun way to play with the feedback concepts.

Also, there are life simulation games based on genome and neural networks.

http://en.wikipedia.org/wiki/Polyworld

I've mucked around with carver mead style neural networks (using a spice simulator) but never with the hopfield style networks.

Robots is very primitive, more primitive than worm and working with programed software and not with neural networks. There is some robots which working on neural networks, but need more such experiment to prove that they working not like programed robot. http://vesicle.nsi.edu/nomad/darwinvii.html

QUOTE

We already have humans, computers, and robots.