If Neural Network was possible to simulate human mind

If Neural Network was possible to simulate human mind...
What will that look like?
A cuboid neural network, this is how 65000 look like
https://www.youtube.com/watch?v=25TwDdUvWwU&feature=channel

How much neural netwoek neurones would it require to simulate a person?
10 billion+ that's more than 13 degits.

What kind of hardware will it need to get 10 billion neurones?
As current GPGPU has Massively Parallel Processors to compute anything faster than CPU do.

However to simulate 10 billion how much memory does it needs?
If it were made up of pixels, the roughness of each neurones/pixel would be the amount of bits
Lets say we have 1024 x 1024 x 1024 in size, if it was compute in 16 bits, the amount of memory requite would be 16*((1024^3)/8 ) = 2147483648 bytes which is 2GB exactly. That was only 1,073,741,824 ,neurones the 1/1000 of what we need.

So I wrote a little calculator in flash which does that calculation above in bits and pixel cube




I would say 8 bit to 12 bit is good enough for any data come to mind, 12 bits is good enough for colour.
if we set cube width to 50000 and roughly 8 bit, a GTX 480's memory bandwidth still wont withstand the memory rate either. Not any harddrive will able to contain 1.25x10 to power of 14 bytes .
It is logically impossible in 10 years times, probably couple of <1nm chip GPGPUs will able to scrape it.

With the current cpu power.... assuming cpu double every 3 years.
It will take about another 25 years or so.
Yes, technology is moving very slow.

Aiko wrote:With the current cpu power.... assuming cpu double every 3 years.
It will take about another 25 years or so.
Yes, technology is moving very slow.

I think it is still possible, although with some compression algorithm to save up memory and use high speed storage like SSD for virtual memory. Although I think it will run very slow

maybe with a super computer it will work
as AIKO said it will take time

spring wrote:maybe with a super computer it will work
as AIKO said it will take time

they got some supercomputer that is trying-currently it can only simulate less than 1000 neurons and eats up over $1,000,000 US in energy costs.
gotta find the article on it soon then.

if you want to use a super computer in an robot like aiko
its impossible the size is too much i think
may a robot that have a size of building is good for super computer
and dont forget the nuclear reactor inside

Discoman wrote:

spring wrote:maybe with a super computer it will work
as AIKO said it will take time

they got some supercomputer that is trying-currently it can only simulate less than 1000 neurons and eats up over $1,000,000 US in energy costs.
gotta find the article on it soon then.

is that really 1000 neurones? That must be a very precise simulation (many weight and value per neuron), rather than a rough one.
I hope you'd find that article, I hope it is quite outdated !

mikeccuk2005 wrote:is that really 1000 neurones? That must be a very precise simulation (many weight and value per neuron), rather than a rough one.
I hope you'd find that article, I hope it is quite outdated !

1.6 billion neurons:
http://blogs.discovermagazine.com/80beats/2009/11/20/ibms-billion-neuron-simulation-can-match-a-cats-brainpower/
http://www.popularmechanics.com/technology/engineering/extreme-machines/4337190
http://arstechnica.com/science/news/2009/11/ibm-makes-supercomputer-significantly-smarter-than-cat.ars
(http://www.technovelgy.com/ct/Science-Fiction-News.asp?NewsNum=1021)

Well, I have done some neural network calculations, because I am designing an FPGA based "nerve center" for a hobby humanoid project. If you are trying to simulate the brain, then one of the main questions is how precise you want it to be. The brain power far exceeds even the simplest simulations on dedicated high cost hardware. Also the simulation is not enough, because you need something more for it to gain consciousness. Nobody knows where it comes from, but I suspect it has something to do with quantum fluctuations. This is a wild thought and just about as far as currently known physics can go. But consciousness may as well be somewhere else, we don't know.

Ok, here are some numbers. If I try to simulate the brain roughly how I think it works directly in hardware (not in software as it is usually done when you have regular CPU), then let's say having a 200MHz FPGA with hundreds of specialized cores with 1GB of RAM, then you can't do better than being an order of 1000000 times "less powerful" than a regular human brain. Fortunately in the area of visioning I think it is possible to get as close as 1000 times "less powerful" than brain with similar approach.

Lets try to put these numbers on scale. Assume you have 1000000 such specialized FPGAs and 1GB of RAM for each totaling 1 petabyte. Also assume that there are no other components in the system. When each FPGA+RAM uses lets say 5W then you need 5 megawatts to power it. Basically you need a power plant for this monster.

Bad news is also that chips can't get smaller indefinitely. Currently the best commercially available technology out there is 28nm. I am sure they can get it to about as small as 10nm, but not much smaller. The problem is that quantum fluctuations will ruin the steady state chips need to function. The good thing is that when consciousness really lays there, we might be able to build a real "artificial intelligence", although it will be quite stupid compared to human.

Consciousness is just an illusion like feelings and all the stuff brain creates. I.e. there are no colors you see colors because it's an illusion made by the brain - like there's no red or blue they are just wavelengths. And there's no pain it's an illusion by the brain.

ico wrote:Consciousness is just an illusion like feelings and all the stuff brain creates. I.e. there are no colors you see colors because it's an illusion made by the brain - like there's no red or blue they are just wavelengths. And there's no pain it's an illusion by the brain.

You're right about that, but I was more keeping in mind self-consciousness, the very thing that makes you feel alive. Even when you know all the electrochemical reactions in the brain and body you still can't answer why you, as an observer, sit in that body. But the fact is, self-consciousness is inherent in the nature. Imagine piece of rock, burning flames, sunflower, earthworm (having one of the most basic neural systems), mouse, human. Somehow the self-consciousness seems to build up as things get more complex, but it doesn't happen to computers and algorithms. As long as we "program" the AI, the nature can't do its work and make things self-aware, if I may say so. If we could integrate "the algorithm of the nature" into some algorithm that our PCs (or as well as supercomputers) can run today, we might get pure non-organic life that may exist in many forms unimaginable today. I don't know about others, but I am waiting that day and might as well contribute to research on this field as well as I can, because I believe that in the future this advancement is inevitable.

But what's self-consciousness ? That you know you are you and others aren't you ? Watch N.G.E. ep 25,26. You know that you are you because you see others and you know they are not you. For example if you see a rock you know that the rock isn't you because you can for example move according your will but the rock doesn't move just because you want it, you have to move it by yourself. Or in a mirror you know that it's you because it acts the same way you act and it reflect all things around you so is logical the person it reflects is you. To create such self-consciousness it only needs simple logic, simple intelligence.

Or do you mean self-consciousness had you think in your head ? - that's not different when a PC computes data in the RAM. The PC has just no intelligence so he can't know what data he computes. If it can associate the data with other data it can create self-consciousness.

The most important thing is the intelligence itself - the logic, recognizing of object, learning etc - and with that comes the consciousness and self-consciousness. Consciousness is created from intelligence so the more intelligent a machine will be the more developed consciousness it will create/have.

It's all only an illusion.

Just an Idea. Maybe possible in a few years or so to just simulate 2,000,000,000 neural net on a GPU at a very rough level ie 8bit per neuron (for reasons that ASCII characters to binary is 8 digits mostly); with another system that controls the robot's hardware and movement and can only influence by the neural network. The two AI's, neural net and main system send and receiving text data of what happen and what to do. The neural net is actually not know how to walk or recognise things, but the main system does(ie BRAINS software). I hope you get my idea ..

mikeccuk2005 wrote:Just an Idea. Maybe possible in a few years or so to just simulate 2,000,000,000 neural net on a GPU at a very rough level ie 8bit per neuron (for reasons that ASCII characters to binary is 8 digits mostly); with another system that controls the robot's hardware and movement and can only influence by the neural network. The two AI's, neural net and main system send and receiving text data of what happen and what to do. The neural net is actually not know how to walk or recognise things, but the main system does(ie BRAINS software). I hope you get my idea ..

sorta like heirarchy PLCs.
the main brain controls the emotions, language, etc. it tells the sub-brains what to do.
the sub-brains directly control the movements. they receive a command like walk forward, and they interpret it into motor controls, resulting in forward motion.
you guys know what I mean, right?
pros: saves processor space and power, allows walking program to be edited without having to edit as alrge of a program
cons: more processors to vent heat from, more expensive

Discoman wrote:

mikeccuk2005 wrote:Just an Idea. Maybe possible in a few years or so to just simulate 2,000,000,000 neural net on a GPU at a very rough level ie 8bit per neuron (for reasons that ASCII characters to binary is 8 digits mostly); with another system that controls the robot's hardware and movement and can only influence by the neural network. The two AI's, neural net and main system send and receiving text data of what happen and what to do. The neural net is actually not know how to walk or recognise things, but the main system does(ie BRAINS software). I hope you get my idea ..

sorta like heirarchy PLCs.
the main brain controls the emotions, language, etc. it tells the sub-brains what to do.
the sub-brains directly control the movements. they receive a command like walk forward, and they interpret it into motor controls, resulting in forward motion.
you guys know what I mean, right?
pros: saves processor space and power, allows walking program to be edited without having to edit as alrge of a program
cons: more processors to vent heat from, more expensive

That is Exactly what I mean. Heat ventilation in this processing scale can be done by water-cooling, would still reach around 50-60C. if the radiators do not get enough air flow (like from breathing air from her mouth/nose).

The very downside is programming the sub-brain to do all kind of things. However this is easier to add a new program to it rather than having a large main neural net brain that needs real training, which the result may not be successful either.

If teaching movement made alot of trouble for programming, 2 neural net brain and one sub brains that commutate will benefit; here's how:

Neural net #1 controls the emotions, language, etc. it tells the sub-brains what to do.
The software(sub brain) translate the language and control the physical body, talk, walk, wave hand. Output to Neural net#1 on what happen ie "I am holding a pen on a piece of paper", "System: you have learned shake hands".
Neural net #2 receive input for intimate learning, how to move, tonal voice, synapse reaction. They become Learning sequence are saved to the sub-brain. This neural net is smaller and more rough calculation ie4 bit the 64 angles the muscles can move (thus this will be correct by the sub-brain make it smoother etc)

There are many different software Aid could possibly do to make a small virtual illusionary being smart !

I found some PDF papers with interesting information on spiking neural network.
http://www.hicomb.org/papers/HICOMB2010-02.pdf

They tested with many processors and CPUs, optimisations.
They tried to prove that neuron over 2.8m with tradition methods will have decline in performance, which makes GPU, Xeon, AMD working around the same performance.
With Hodgkin–Huxley and Fast Math, performance jumped 2-3x
Some simulation are as large as 6m neurones

They also mentioned that 2.8 million neurons takes around 581 MB, so 4820 neurons per MB.
Today's 1GB GPU like low profile gts 450 would able to do almost 5m. Now that is insane - consider that you can still add some share memory from the CPU over PCIe. PCIe v2 bandwidth is 8GB/s.