Human level AI
Post Reply
 
Thread Rating:
  • 0 Votes - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
09-11-2016, 02:46 AM
Human level AI
I've seen a number of mentions of this in other threads so I assume it's a topic of interest.

The hardware requirements for this class of machine are quite different from classical AI, scientific, or engineering applications. As such, there is no hardware available today that could be used to build an efficient system of this type. You could build one by brute force, but the cost with current technology would be $10-20 billion.

You could probably get this down into the mainframe range of $20 million in five or six years. You could probably also reduce this to minicomputer prices of perhaps $2 million within ten years. I'm not a chip designer so it's difficult for me to say if a consumer level price is possible.
Find all posts by this user
Like Post Quote this message in a reply
09-11-2016, 11:39 AM
RE: Human level AI
TrueNorth® chip.













#sigh
Find all posts by this user
Like Post Quote this message in a reply
[+] 2 users Like GirlyMan's post
11-11-2016, 11:03 AM
RE: Human level AI
With only one response, this thread is likely to be short. I had thought this was a topic of interest; it seems not. The response was brief and without questions. This perhaps suggests confidence. I'm not sure where the confidence would come from other than maybe the length of time that IBM has been involved with computing.

Not mentioned was the Stanford Neurogrid board.

Neurogrid simulates a million neurons connected by billions of synapses in real-time, rivaling a supercomputer while consuming a 100,000 times less energy—five watts instead of a megawatt!

Not mentioned was Brain Scales.

Each 20-cm-diameter silicon wafer contains 384 chips, each of which implements 128,000 synapses and up to 512 spiking neurons. This gives a total of around 200,000 neurons and 49 million synapses per wafer.

Back in 2009, Markram gave a TED talk where he stated that a replica of a human brain could be built in 10 years. The above chip is the product of the FACETS project which ran under Markram from 2011-2015 using a Blue Gene computer as a workhorse system. There was a proposal to add the neural boards to the JUGENE computer much like was done with Cray's Oak Ridge Jaguar computer. In 2013, the project was funded into a second, 10-year phase with $1.3 billion. This project had essentially zero chance of success. As Christof Koch said, “The roundworm has exactly 302 neurons, and we still have no frigging idea how this animal works.”
They began to realize this last year, https://www.scientificamerican.com/artic...to-fix-it/. One recalls Japan's vaunted 5th/6th generation computing project with similar goals.

Also, not mentioned was Nvidia's Tesla P100 which can do 8 TeraFLOPS. This board was added to Jaguar to create Titan at 17.59 petaFLOPS.

It seems logical to conclude that there is an assumption something like:
1. The IBM True North chip is a workable building block for human-like AI.
2. IBM is now developing such a system.

To discuss the suitability of True North I would need the technical specs covering processing speed, data width, memory speed, width, and size. I don't know if this is available. However, I might point out that even if the hardware were adequate, you would still need an architecture. This doesn't exist since there is currently no model for how the brain works. As Christof said, not even for a roundworm.

If the True North chip is that good, one might also wonder why it isn't going into IBM's Summit supercomputer which is estimated to develop 200 petaFLOPS.
Find all posts by this user
Like Post Quote this message in a reply
11-11-2016, 11:42 AM (This post was last modified: 11-11-2016 11:46 AM by Mathilda.)
RE: Human level AI
I'll probably regret chiming in here. I've given up talking about AI on forums or social media recently because it seems people can get really upset if you tell them stuff that ruins their fantasy.


(11-11-2016 11:03 AM)scientious Wrote:  As Christof Koch said, “The roundworm has exactly 302 neurons, and we still have no frigging idea how this animal works.”

This. Christof Koch knows what he is talking about. His book Biophysics of Computation is like a bible of computational neuroscience. In it he describes how complicated just one single neuron is. A biologically plausible neuron is far more complicated than a whole artificial neural network. A single model of a biologically plausible neuron is more than just a integration device with a threshold which most neural networks use.

Yeah having a chip that can process a load of neurons is nice, but it's only a start. Because even if we were able to process all the 100 billion neurons and quadrillion synapses of a human brain, we still wouldn't have enough processing power to create human level intelligence.

You'd then have to have several orders more magnitude of processing power in order to find out how to use them. This is because you don't just need to build a brain but evolve it as well. With any species, a population has existed over many thousands of generations, with each mutation or configuration tested multiple times in real world environments. Mechanical movement is extremely slow compared to computers and all intelligence needs to be embodied in an environment of some sorts.

But the main limiting factor is our own ability to actually understand and engineer neural networks for artificial general intelligence. Even if we had as much processing power as we could ever want we'd still need to understand it all.

Add to that the ability to measure at a fine enough scale, research into batteries, sensors, actuators and materials etc. Processing power alone is not enough.

Personally I'm not even sure that we'll be using neural networks so much in the future once we understand a lot more how they actually work. There are other techniques out there that work just as well with greater explanatory power and therefore should be easier to engineer.
Find all posts by this user
Like Post Quote this message in a reply
11-11-2016, 12:47 PM
RE: Human level AI
(11-11-2016 11:42 AM)Mathilda Wrote:  I'll probably regret chiming in here. I've given up talking about AI on forums or social media recently because it seems people can get really upset if you tell them stuff that ruins their fantasy.
Well, I will probably regret this too. There is ongoing work in this area, but I am not free to discuss it because it isn't published. The latest estimate was publication in 24 months. This may no longer be valid.

Publication would cover the evolutionary development of human cognition, a detailed description of human cognition, explanations of what architecture would be involved with duplicating it, and some miscellaneous discussions of religious and social implications. The publication isn't likely to make very many people happy. It gives a strong refutation of the notion that a mind can exist outside of a brain or that a mind can exist in isolation (which refutes dualism in any form). It shows limitations of the mind and that the brain operates primarily to avoid overloading the higher order functions (which refutes motivational speakers who claim unlimited potential). It gives equations on what conditions are necessary for cognition to work based on knowledge theory. It gives solutions to classical problems like the Frame problem but also covers new problems like the origin of abstractions (which refutes William Lane Craig) or abstract valuation (which refutes Harris' view of free will). It even shows the basis of prejudice and creativity.

So, when I talk about hardware it isn't groping in the dark. The theory provides the architecture. If you have appropriate hardware and software and you put in the time, money, and effort I don't know of anything that would prevent it from working. This would create human-like cognition in a computer based system. Would such a machine be conscious? I have no idea, but it would act as though it were.

What are the likely responses now:
1. You are crazy, prove you have a theory.
This is pointless; there is no proof prior to publication. And, even then it would take a lot more work to prove the theory and prove that a system was possible.

2. You can't come up with something that IBM, Google, Cray, etc. haven't. These projects are funded with hundreds of millions of dollars.
The money has been for hardware and some AI development. There hasn't been any real effort put into a cognitive system since there is no theory. However, once the basic theory was known, I'm sure there would a massive effort to prove/disprove it, refine it, expand it, etc. I'm certain that subsequent research and development would go beyond anything the original theory covered.

3. If you could explain the evolutionary development of the mind then someone would have done it already.
I think that is the nature of the problem. No one seems to have been able to create a theory of consciousness from a purely philosophical point of view; the closest has probably been the Cyc system. No one has been able to explain it from a purely AI point of view; I suppose Watson would be the closest. Behavioral psychology covers what rather than why or how. And then you still need knowledge theory. I doubt enough information was available before now.

4. You can't be right because I know that I am psychic.
Can you could build a psychic computer so that it could be studied? I believe a cognitive computer could be built.
Find all posts by this user
Like Post Quote this message in a reply
11-11-2016, 12:55 PM
RE: Human level AI
(11-11-2016 12:47 PM)scientious Wrote:  If you have appropriate hardware and software and you put in the time, money, and effort I don't know of anything that would prevent it from working.

Right now? NP.

#sigh
Find all posts by this user
Like Post Quote this message in a reply
11-11-2016, 12:58 PM (This post was last modified: 11-11-2016 01:33 PM by GirlyMan.)
RE: Human level AI
(11-11-2016 12:47 PM)scientious Wrote:  I believe a cognitive computer could be built.

I believe something that for all intents and purposes appears capable of cognition can be built. Is it gonna work like our wetware? Not anytime soon. But someday. Fortunately for us any function can be approximated to an arbitrary degree with a mutlilayer perceptron. My point with the IBM post is that we now have the hardware to efficiently find approximations for complex problems we could not approximate before. Hence the recent resurgence in ANN research now that the training which would have taken months back in the day can be done in seconds. While a computational model of cognition would be extremely useful I don't think it is required. A sufficient approximation is enough.

#sigh
Find all posts by this user
Like Post Quote this message in a reply
11-11-2016, 02:55 PM
RE: Human level AI
(11-11-2016 12:47 PM)scientious Wrote:  Well, I will probably regret this too. There is ongoing work in this area, but I am not free to discuss it because it isn't published. The latest estimate was publication in 24 months. This may no longer be valid.

Publication would cover the evolutionary development of human cognition, a detailed description of human cognition, explanations of what architecture would be involved with duplicating it, and some miscellaneous discussions of religious and social implications. The publication isn't likely to make very many people happy. It gives a strong refutation of the notion that a mind can exist outside of a brain or that a mind can exist in isolation (which refutes dualism in any form). It shows limitations of the mind and that the brain operates primarily to avoid overloading the higher order functions (which refutes motivational speakers who claim unlimited potential). It gives equations on what conditions are necessary for cognition to work based on knowledge theory. It gives solutions to classical problems like the Frame problem but also covers new problems like the origin of abstractions (which refutes William Lane Craig) or abstract valuation (which refutes Harris' view of free will). It even shows the basis of prejudice and creativity.

So, when I talk about hardware it isn't groping in the dark. The theory provides the architecture. If you have appropriate hardware and software and you put in the time, money, and effort I don't know of anything that would prevent it from working. This would create human-like cognition in a computer based system. Would such a machine be conscious? I have no idea, but it would act as though it were.

Sounds very interesting. I would love to read it when it is published. If you could send me a PM when it is then I'd appreciate it.

Thanks
Find all posts by this user
Like Post Quote this message in a reply
Post Reply
Forum Jump: