SPAUN largest human brain model yet - Nengo Open Source Neural Simulator

[kram1032]

Really interesting. We're getting somewhere with neural network AI 
http://nengo.ca/

[eiffie]

I really like AI and neural networks but when I hear "SPAUN's errors are similar to human errors" I begin to appreciate the bit shifting awesome dude giant calculator I call my laptop.

[kram1032]

lol.
Well, it's meant to be a model of human mind. It's still extremely limited. What they did is essentially modelling the most important areas for cognitive tasks of the human brain, reducing the neuron count a lot and making it static so it won't actually learn something new.
So there are other neural nets that can actually learn new stuff but aren't as well-developed.
Making it actually form new connections and neurons is the next step.

[cbuchner1]

Really interesting. We're getting somewhere with neural network AI 
http://nengo.ca/

Interesting, they even have some partial support for GPU acceleration with CUDA.
http://ctnsrv.uwaterloo.ca/docs/html/advanced/gpu.html

Christian

[cKleinhuis]

the question is if building a computational neuronal network would be able to produce non-deterministic output, since we learned in informatics that the computer is fully deterministic and thus would not be able to help other computers to solve their problems, it all comes down to the available input, e.g. how can a computer even check if another computer is working right when it does not have the possibility to press the on/off button ... it might produce some unexpected results though ... and if at the end this computational neural network could think ... this would bring us to the philosophical question if our brains are deterministic at the end ... or better said, that pure deterministic approaches can lead to undeterministic results...

[kram1032]

For now it's accepted as a fact, that human brains are turing complete and not more. In other words, what ever we can think of, a computer can think of too. It's just a matter of a corresponding program. Furthermore, most evidence so far hints at the fact that the brain does NOT utilize quantum effects and thus "sufficiently" deterministic as well.
And if you want true non-determinism to be introduced into the system, you could always get yourself a quantum random number generator and use that to introduce noise into the system.
The computer's main advantages are that it doesn't (almost ever) make errors (the errors almost always stem from programmers or users) and that it's much faster.
Both those advantages are pretty much gone in this system since it's prone to errors and the many many neurons that need to be simulated are relatively slow to simulate.
If you take deterministic as non-chaotic, that's nonsense because if that were the case, a computer could never ever render any kind of fractal.

Note that the current implementation, as I already mentioned, is very fixed and rigid. It can only recognize A, P, K,  [, ] and 0-9 as fixed letters as well as 0-9 as hand-written letters.
Any other input is ignored... or misinterpreted as falling into one of those categories. It also can't learn anything. Even the letter recognition, as far as I can tell, is currently fixed through offline-learning. While it performs tasks, essentially while it's "on", it can't learn anything new. Even when in persumably the off-line learning mode, it can only work within the categories it already knows.

All that being said, learning has been tackled plentifully in other resarch neural networks so it's probably not all that difficult to upgrade this architecture to include state-of-the-art online learning and an expanded scope of understanding more varied input.
And the things it already can do, like inductively guessing patterns or understanding input AND copying the input style, are quite amazing features.

Determinism doesn't exclude chaos. About 99% of this website's content proves that over and over. And about 99% of those 99% are generated by computers.

[Alef]

Matrix has you