Artificial neural network

[3DModelerMan]

What would need to happen for someone to write an artificial neural network that could gain sentience? Would you need speaker outputs, a camera for "eyes", and microphones as the inputs? And then train it? If it was massively parallelized then you could run an equal number of neurons per processer core. How many cores would it take and at what clock speed? Like, if someone wrote with creative enough optimizations, would it be possible for a desktop with 6 cores at 3.3Ghz and 8-16 GBs of RAM to gain sentience? And what about GPGPU programming?

[kazymjir]

I was thinking about this one time... on acid, lol.
My idea was to remove radiator from CPU, put one LSD blotter on CPU, and then mount radiator again.
I was thinking that CPU will get consciousness of his existence in this world... but when I looked at my PC, and on my blotter, I decided to save this blotter for another time, lol.

[CuteAlien]

We don't really know yet how sentience works, so rather tricky question :-)

If you want a really, really long answer about what we might need to emulate brains, here is a technical report about this: http://www.fhi.ox.ac.uk/__data/assets/p ... report.pdf
Some people are actually working on this in the Blue Brain Project: http://en.wikipedia.org/wiki/Blue_Brain

Concrete numbers vary a lot, but our current super-computer generation is approaching the very lowest estimates I've seen so far for the needed calculating power to reach human-like brain capacities, but more should get there this decade. Which means you probably must wait another 10-20 years to have that power on your desktop (or maybe on your mobile phone by then).

Then again this is about the computing power of the brain - we don't know when organisms reach sentience (we don't even know for sure if it is a function of the brain although nearly everyone these days suspects so except the religious crowd, some hippies and maybe a few mystics). So it is entirely possible that your computer already is sentient, but just has no way of telling you.

[Granyte]

i don't think it's related to computing power as much as beeing related to the desing of the neural network

i'm not even sure we have a neural network that can learn with it self without an external rule telling him the upgrades to do

[3DModelerMan]

I haven't really heard that much about artificial neural networks lately either. Is there not much research being done on the subject?

[CuteAlien]

There is a lot of research going on in natural neural networks (aka brains). Mainly people still trying to figure out firing patterns, and which parts of that actually do matter. For example we learned in recent years that firing rhythm is also a pattern to transport information - and not just the spikes themselves. Current work is a lot about synchronizing artificial networks with natural networks to get similar results (there are some very simple brains in certain snails for example which are great for testing that stuff). And a lot of discussion is going on on which level the neurons should be simulated - do you for example also need to simulate neurotransmitters and their distribution in the brain?

As for learning without external rules - that's called "unsupervised" learning. There's a few algorithms for that, but they are often way slower than the supervised rules.

A very big problem with neural networks is that it is very hard to tell what is going on. Figuring out why a certain network works better/worse than any other. Unless that problem is solved I don't expect real breakthroughs in combining them to form larger networks.