[quote name='Hodgman' timestamp='1349537905' post='4987416']
The brain operates at only ~10Hz and yet in terms of NN complexity, it delivers as many FLOPS as our most powerful supercomputers -- though they require several megawatts and fill large rooms, whereas the brain only needs ~20 watts and fits in your head. So not only is it ridiculously more powerful than a regular computer, it's also ridiculously small and energy efficient.
Yes, this is due to it's specialized design -- to compete with it, you would have to build your fake brain using similar construction processes (if you didn't want you brain to fill a room and require more rooms of power equipment). If you search for "neural network hardware", you'll find this isn't a new topic. People have built hardware ANN's before that were as complex as a cat's brain, though with not very impressive results.
Keep in mind that it's easy to get caught up with the state of technology and think that we're an advanced race, but biological construction techniques really do put our technology to shame. We're still playing with sticks and fire compared to what biology is doing. It's going to be a long time until we can build hardware that is similar to a brain.
e.g.
* Plants are mostly made out of air, are self constructing, can be endlessly recycled and contain photosynthesis units that are absolutely amazing compared to our best solar cells. That's why we use plants to make biofuel, or dig up million year old plants to burn as oil.
* Kevlar is produced by stretching carbon nanotubes through a pressurized vat of extremely hot concentrated sulfuric acid (which is dangerous and hazardous to the environment). Abalone shells are tougher than kevlar, and are made out of seawater and sunlight, produced in a pollution-free process on demand.
Yeah the power consumption would be incredible. What about a molecular computer or a quantum computer?
I still think the modern computer model, as useful to do math as it is, is pretty outdated for the things we want it to do. Fetching each instruction at a time with millions of logic gates, which need constant voltage, takes up a lot of power and time. Not sure if anyone has made a different/more efficient kind of computer. I would imagine an analog computer to be faster but because the voltage determines the number instead of a binary number that means that there could be more power consumption, unless you do decimal voltage.
[/quote]
Binary works even if there are fluctuations in the currents/voltages so we can make the processor smaller. Id imagine youd need all kinds of fancy components and regulation systems to keep the currents/whichever it is stable.
[quote name='CryoGenesis' timestamp='1349540408' post='4987427']
[quote name='Hodgman' timestamp='1349537905' post='4987416']
The brain operates at only ~10Hz and yet in terms of NN complexity, it delivers as many FLOPS as our most powerful supercomputers -- though they require several megawatts and fill large rooms, whereas the brain only needs ~20 watts and fits in your head. So not only is it ridiculously more powerful than a regular computer, it's also ridiculously small and energy efficient.
Yes, this is due to it's specialized design -- to compete with it, you would have to build your fake brain using similar construction processes (if you didn't want you brain to fill a room and require more rooms of power equipment). If you search for "neural network hardware", you'll find this isn't a new topic. People have built hardware ANN's before that were as complex as a cat's brain, though with not very impressive results.
Keep in mind that it's easy to get caught up with the state of technology and think that we're an advanced race, but biological construction techniques really do put our technology to shame. We're still playing with sticks and fire compared to what biology is doing. It's going to be a long time until we can build hardware that is similar to a brain.
e.g.
* Plants are mostly made out of air, are self constructing, can be endlessly recycled and contain photosynthesis units that are absolutely amazing compared to our best solar cells. That's why we use plants to make biofuel, or dig up million year old plants to burn as oil.
* Kevlar is produced by stretching carbon nanotubes through a pressurized vat of extremely hot concentrated sulfuric acid (which is dangerous and hazardous to the environment). Abalone shells are tougher than kevlar, and are made out of seawater and sunlight, produced in a pollution-free process on demand.
Yeah the power consumption would be incredible. What about a molecular computer or a quantum computer?
I still think the modern computer model, as useful to do math as it is, is pretty outdated for the things we want it to do. Fetching each instruction at a time with millions of logic gates, which need constant voltage, takes up a lot of power and time. Not sure if anyone has made a different/more efficient kind of computer. I would imagine an analog computer to be faster but because the voltage determines the number instead of a binary number that means that there could be more power consumption, unless you do decimal voltage.
[/quote]
Binary works even if there are fluctuations in the currents/voltages so we can make the processor smaller. Id imagine youd need all kinds of fancy components and regulation systems to keep the currents/whichever it is stable.
[/quote]
I don't think you would if you built it properly. The main problem would be the heat. Also, if any of the wires touched it would just melt the entire computer (but I think that happens with binary processors too).
Indeed. It's a waste of time to even try to write a computer program that gives the computer consciousness. John Searle put that to bed more than 30 years ago.
The term "artificial intelligence" as it is used today is a bit of a misnomer. I can see why it would cause some confusion.
I don't think talking about conciousness is relevant. Anything that looks like it has conciousness is enough. I can only be sure about my conciousness. I think the Duck test applies here well.
Alan Turing was many things (genius, spy, homosexual), but I think nobody would call him a duck.
Stephen M. Webb
Professional Free Software Developer
[quote name='szecs' timestamp='1349536832' post='4987412']
I don't think talking about conciousness is relevant. Anything that looks like it has conciousness is enough. I can only be sure about my conciousness. I think the Duck test applies here well.
Alan Turing was many things (genius, spy, homosexual), but I think nobody would call him a duck.
[/quote]
Ahh. teh illiterateness of me.
Anyhoo, I didn't call him a duck, I called the computer a duck. Even Wiki has an article about this whole conciousness, can we forget the question of "can machines think" thing already?
I wish I knew enough to know how far you can simplify a neuron down to and still have something functional.
On the up side though you can biologically cheat by allowing some information to always be free, IE accessible where as that information would need to be propagated neuron to neuron in nature.
The hardware requirements aside (a field I know little about), I think people expect an AI to passively sit there being fed data and then kapow it's alive. Our intelligence (and that of animals) is very embodied. I think that our body is what allows us to get the continual feedback that high quality learning requires. Every time we move a muscle it affects our whole body via physics, which affects all our senses, so therefore there's a very tight and high bandwidth learning loop. Now I'm not saying that there aren't other good paradigms, but we know this one works.
Conversely our passive neural nets make screwball mistakes like learning from satellite photos that a tank is present if it's a sunny day because all it's tank training data was taken on a sunny day. I'm not surprised. It's getting one bit of feedback (present, not present). It really knows nothing about the properties of tanks or how you interact with tanks or why they matter. Light vs dark was the easiest neural net that matched the training data.
What makes you think that doing things in parallel is the key to consciousness?
There's also the point that, as far as running an algorithm is concerned, there is no difference whether things are run in parallel or not - a single core machine can always do the same thing as one doing it in parallel, it's just a question of performance. If consciousness is something that can be developed on a computer (as we know it today, i.e., simply a matter of running the right software), then it can be run on any computer, although it might be very slow.
As far as algorithms are concerned, what you suggest is nothing new - e.g., neural networks are used in AI, and they model a process that occurs in parallel. But it doesn't matter whether you run it on hardware with multiple cores or only one.
You also seem to be unaware of anything to do with multiple cores, SMP and so on? - Yes, writing things to run in parallel is a difficult area, but this is not specific to AI, and is something that is already done today on most computers.
In summary: (a) running things in parallel isn't inherently necessary, it's just a question of whether you'd get better performance, (b) the industry already realised years ago that running things in parallel is a way to get better performance, and this transition became standard even on bog standard PCs in the last decade - even phones have multiple cores these days. (Of course, it's true that we're still a long way from solving the problem of scaling most software to say thousands or millions of cores, but again, it's not like no one's thought of this.)
What makes you think that doing things in parallel is the key to consciousness?
There's also the point that, as far as running an algorithm is concerned, there is no difference whether things are run in parallel or not - a single core machine can always do the same thing as one doing it in parallel, it's just a question of performance. If consciousness is something that can be developed on a computer (as we know it today, i.e., simply a matter of running the right software), then it can be run on any computer, although it might be very slow.
As far as algorithms are concerned, what you suggest is nothing new - e.g., neural networks are used in AI, and they model a process that occurs in parallel. But it doesn't matter whether you run it on hardware with multiple cores or only one.
You also seem to be unaware of anything to do with multiple cores, SMP and so on? - Yes, writing things to run in parallel is a difficult area, but this is not specific to AI, and is something that is already done today on most computers.
In summary: (a) running things in parallel isn't inherently necessary, it's just a question of whether you'd get better performance, (b) the industry already realised years ago that running things in parallel is a way to get better performance, and this transition became standard even on bog standard PCs in the last decade - even phones have multiple cores these days. (Of course, it's true that we're still a long way from solving the problem of scaling most software to say thousands or millions of cores, but again, it's not like no one's thought of this.)
It's less about parallel processing or multiple cores it's more about nodes. The reason you would want to make the nodes hardware instead of digital software is because to update all the nodes would take time but if it were hardware they would all be updating themselves constantly taking in inputs, making decisions and spitting outputs. Almost as if you had a silicon neuron and the wires were silicon synapses. I know another thing that would be impressive, the nodes could replicate themselves (somehow).
But, I honestly think that conciousness isn't as complicated as people are suggesting. Sure the whole thing in itself is complicated but the little systems that make it work aren't at all. I think of it as more of an illusion than anything else.
It's less about parallel processing or multiple cores it's more about nodes. The reason you would want to make the nodes hardware instead of digital software is because to update all the nodes would take time
Yes, this is what I say - it may be slower, but that doesn't change whether you can do it. No doubt any computer capable of human level consciousness would have to be massively parallel, but that's separate to whatever the key to consciousness might be. If you can run it on a massively parallel computer, then you could do it on a non-parallel one too, albeit slower.
Also I don't see how multiple cores is different to "nodes" - multiple cores are doing parallel processing in hardware, for real, and not in "software".
But, I honestly think that conciousness isn't as complicated as people are suggesting. Sure the whole thing in itself is complicated but the little systems that make it work aren't at all. I think of it as more of an illusion than anything else.[/quote]Which people do you mean suggest this? It's a not uncommon view that consciousness arises out of the complexity of a large number of smaller simpler parts.
