A.I Concept built after Neurons in the brain

[mooeypoo]

I just finished watching the movie "In Its Image". My brain hurts..

 

How plausible to you guys think it is that we will have TRULY thinking machines based on this concept? It sounds mindboggling but beyond that, Im a bit surprised not to have heard of this concept before.

 

Maybe it's just me... anyone heard of this?

 

btw.. the entire thing kinda smells weird... the "lots of products were invented by creativity machines but i'm obliged by contract to not tell you which" is.. uhh.. suspicious, to say the least.

 

Is this valid!?

 

~moo

[Mr Skeptic]

The idea of neural networks is to simulate what happens in neurons. They are a type of generic solution that can be trained to do stuff (eg distinguish between things in different categories). So far, they are mostly simplistic things with three layers of "neurons" and require hundreds of repetitions to learn things. If they made a more complex arrangement and improved on the learning functionality, it is possible that they could form a humanlike intelligence. However, I understand that they currently have about the same intelligence potential as a cockroach.

[mooeypoo]

This movie claims they have made more than just progress -- it claims it has functional machines like that, capable of creating 'creative' processes...

 

exaggeration?

[bascule]

Computational biology offers the promise of realistic models of cellular systems. Projects like Blue Gene and Folding@home are trying to brute-force a lookup table for how proteins fold. Once this is complete, computers will also be much faster, and it will be realistic to start building accurate computer models of cellular and multicellular behavior.

 

Once you have that, you're on your way to artificial organisms, including artificial humans, modeled down to the cellular level. Thus far there's no reason to believe that an artificial human with an artificial brain wouldn't be an intelligent, conscious being.

 

This movie claims they have made more than just progress -- it claims it has functional machines like that, capable of creating 'creative' processes...

 

exaggeration?

 

There's plenty of systems which have emergent "creative" properties. However almost none of them are modeled after the behavior of the high level structures of the human brain. Nor are such systems generally capable of general intelligence.

 

The closest thing to a general intelligence system based off human neurophysiology is Numenta's Hierarchical Temporal Memory. This system is not only capable of creativity but also prediction, and unlike most other "creative" systems is capable of general intelligence.

[iNow]

How plausible to you guys think it is that we will have TRULY thinking machines based on this concept?

 

Very. Maybe not this specific concept, but I'm confident that this type of idea will play a clear role in whatever thinking machines we ultimately arrive with.

 

Thanks for the link. To me, what it "smells" a bit like is a sales promo or funding request to people in the Department of Defense. It raises some interesting thoughts and describes one pretty cool approach. I guess all I can say is, "We'll see."

[theCPE]

Numenta hasn't provided anything exciting yet....

 

Their "research release" is not very useful, and nothing new can be solved with their system that cannot be solved otherwise. Also, pattern recognition isn't a new concept, HTMs are only a slight modification of prior pattern recognition techniques.

 

As usual it is premature to propose the future of AI is here or if truly thinking machines are even possible.

 

What I find really amusing is the multitude of numenta fanboys in their forums that don't even understand or know Hawkins's theories.

[bascule]

Their "research release" is not very useful, and nothing new can be solved with their system that cannot be solved otherwise. Also, pattern recognition isn't a new concept, HTMs are only a slight modification of prior pattern recognition techniques.

 

Harp on the research release all you want. It does include modifications of previous pattern recognition techniques, but the final version most certainly will not. They're using existing algorithms to patch in unimplemented functionality.

 

The goal is to reimplement the neocortical column in software. Among other things, this will facilitate what's absent from the research release: time-sensitive sequence prediction.

 

As usual it is premature to propose the future of AI is here or if truly thinking machines are even possible.

 

I'm yet to see a good argument against computationalism... the best I've seen come from the likes of Searle and Penrose, and they're all terrible.

 

What I find really amusing is the multitude of numenta fanboys in their forums that don't even understand or know Hawkins's theories.

 

I've studied neurophysiology intensively and read Hawkins' book On Intelligence, and was happy to see he came to many of the same conclusions as myself, particularly in regard to the role of neocortical columns in consciousness, as well as thalamocortical loops.

[theCPE]

The goal is to reimplement the neocortical column in software. Among other things, this will facilitate what's absent from the research release: time-sensitive sequence prediction.

 

Yes, I am familiar with numenta's goal. But that is my point.

 

Sometimes people forget what HAS been done and what people are ATTEMPTING to do (goal). That is why it is always entertaining to look back at predictions.

 

I'm yet to see a good argument against computationalism... the best I've seen come from the likes of Searle and Penrose, and they're all terrible.

 

Interesting, I haven't heard any good arguments of why it should be possible to create AI.

 

 

I've studied neurophysiology intensively and read Hawkins' book On Intelligence, and was happy to see he came to many of the same conclusions as myself, particularly in regard to the role of neocortical columns in consciousness, as well as thalamocortical loops.

 

Cool, I'm sure Hawkins would be pleased to know someone on scienceforums.net came up with the same theories he did independently.

 

BTW that quote of yours reminds me of the bar scene in Good Will Hunting....you know...the scene where the guy tries to pawn off other people's ideas as his own....

 

I've read his book too. Pretty good book with lots of interesting ideas, definitely makes you ponder. I'm eager to see what success his HTMs can provide but I'm not expecting anything much more sophisticated than pattern recognition within expert systems.

[Mr Skeptic]

Interesting, I haven't heard any good arguments of why it should be possible to create AI.

 

It follows directly from the concept of materialism.

[bascule]

It follows directly from the concept of materialism.

 

Or for that matter, most types of monism. Dualism is really the only metaphysical formulation of consciousness which precludes computationalism, save for weird arguments like Searle's biological naturalism (monistic formulation which stipulates that consciousness can only arise from biological systems) or Penrose's perspective (loosely knit assortment of anti-computationalist arguments centering around the quantum mind hypothesis, asserting what's essentially dualist metaphysics)

[doG]

Jeff Hawkins: Brain science is about to fundamentally change computing is an interesting lecture related to this topic. Perhaps some of you will enjoy it as much as I did.

[theCPE]

It follows directly from the concept of materialism.

 

So anything that exists is made of matter and can be explained.

 

I agree.

 

From there I suppose the leap is that: If something already exists than we should be able to replicate it.

 

That I don't think is necessarily true, just like the silly elementary school saying of "anything is possible" isn't true.

 

Can we create AI or conscientiousness? It might be possible, but I don't think it is a given.

 

And what numenta has released doesn't even approach true AI.

 

Pattern recognition and expert systems can do some incredible things and provide some very sophisticated automated processes but that is still deterministic.

 

It is so annoying when I read articles that were obviously written by people outside of the field predicting silly things like, "By 2010 computers will think like humans" etc....

[Luminal]

So anything that exists is made of matter and can be explained.

 

I agree.

 

From there I suppose the leap is that: If something already exists than we should be able to replicate it.

 

That I don't think is necessarily true, just like the silly elementary school saying of "anything is possible" isn't true.

 

Can we create AI or conscientiousness? It might be possible, but I don't think it is a given.

 

And what numenta has released doesn't even approach true AI.

 

Pattern recognition and expert systems can do some incredible things and provide some very sophisticated automated processes but that is still deterministic.

 

It is so annoying when I read articles that were obviously written by people outside of the field predicting silly things like, "By 2010 computers will think like humans" etc....

 

Well, computers could think like humans given technology at this very moment, if people could understand the human brain well enough to replicate it.

 

Remember, as Hawkins said, human intelligence primarily stems from the neocortex, a relatively small structure compared with other structures such as the cerebellum.

 

Strip from the brain all structures for emotions, desires, sensation, body/organ regulation, motor skills, and you can still have the core of human intelligence.

 

In other words, one needs not replicate the entire human brain to achieve human-level creativity.

 

In addition, the brain is chaotically organized in comparison to a computer. Organization in a computer allows it to have massively repetitive information (such as inheritance) that a human brain cannot achieve.

 

To summarize my post, it is up to programmers to generate AI, not chip developers or supercomputers. We have the power, now we need the finesse.

[bascule]

From there I suppose the leap is that: If something already exists than we should be able to replicate it.

 

That I don't think is necessarily true, just like the silly elementary school saying of "anything is possible" isn't true.

 

Can we create AI or conscientiousness? It might be possible, but I don't think it is a given.

 

The line of reasoning is more like:

 

Consciousness is a byproduct of brain activity (a metaphysical view held by materialists and most monists)

 

The brain is a classical physical system and is thus computable by a universal computer (Searle and Penrose take issue here, and Tegmark attempts to refute them)

 

Therefore, consciousness is a byproduct of computation

 

Clearly neither of the premises of this syllogism are "provable," they are only compatible with the prevailing scientific and philosophical understanding.

 

Those espousing dualism or the brain as non-classical system can argue that computationalism is false since either of these propositions refutes the above premises.

[theCPE]

Well, computers could think like humans given technology at this very moment, if people could understand the human brain well enough to replicate it.

 

Exactly, and that is a large IF.

 

In other words, one needs not replicate the entire human brain to achieve human-level creativity.

 

I agree.

 

To summarize my post, it is up to programmers to generate AI, not chip developers or supercomputers. We have the power, now we need the finesse.

 

Absolutely. Which is why predictions about how "smart" computers will be in X years are so stupid. They rely on the principal of Moore's Law with the idea that computers just aren't FAST enough yet, or something similarly silly.

 

It is up to computer scientist/engineers to develop an architecture or algorithm that models how pattern recognition and association works with neural pathways.

 

My only argument is that discovering HOW and replicating isn't necessarily a given, it could be possible but not necessarily.

 

All things that exist cannot be replicated by us is my assertion.

 

As a simple example, do you think every human being has the ability to create complex algorithms to solve difficult problems?

 

I imagine there is a level at which given infinite time someone could be "dumb" enough to not be able to come up with a correct algorithm to solve some specific problem. Now what if there is a problem or algorithm that needs to be designed that is so complex that it surpasses the capabilities of the smartest humans?

[Luminal]

I have some fairly solid ideas on how to generate human level intelligence in modern computers, and having read up extensively, have not seen similar veins of research taking place in any companies or universities.

 

Hence the whole reason I'm majoring in computer science, to give myself the practical skills to implement such concepts.

 

On the small chance that I've arrived upon something truly unique (I won't get my hopes up ), I try to avoid explaining my ideas online, but suffice to say that researchers are not utilizing genetic algorithms in conjunction with neuroscience half as much as they should.

 

Human intelligence was not programmed; it was evolved.

[ydoaPs]

Human intelligence was not programmed; it was evolved.

And we have computer programs that evolve.

[bascule]

Human intelligence was not programmed; it was evolved.

 

And humans have an ancestry going back billions of years. Evolving strong AI from scratch isn't exactly a feasible idea.

 

The dead simplest approach would be to use computational biology to grow a human from a fertilized zygote inside a computer.

[Luminal]

And humans have an ancestry going back billions of years. Evolving strong AI from scratch isn't exactly a feasible idea.

 

The dead simplest approach would be to use computational biology to grow a human from a fertilized zygote inside a computer.

 

 

It is a feasible idea.

 

1) Generations are arbitrarily short in a computer. Generations in mammals are measured in years, not microseconds.

2) Natural evolution had no guiding hand or designer.

3) Evolution also had no purpose or "means to an end". If something survived, it survived. Within computers, humans can program evolutionary purpose and even sustain generations that are evolving toward something intelligent yet have severe limitations in their current form. How many species that were evolving toward intelligence went extinct because their heads were too large or their brains consumed too much energy? Perhaps dozens; we shall never really know.

4) Most importantly, humans have a highly effective intelligent system on which to model their computers: themselves. Natural evolution had no model to work from.

 

It's a far cry from working from scratch.

[theCPE]

It is a feasible idea.

 

1) Generations are arbitrarily short in a computer. Generations in mammals are measured in years, not microseconds.

2) Natural evolution had no guiding hand or designer.

3) Evolution also had no purpose or "means to an end". If something survived, it survived. Within computers, humans can program evolutionary purpose and even sustain generations that are evolving toward something intelligent yet have severe limitations in their current form. How many species that were evolving toward intelligence went extinct because their heads were too large or their brains consumed too much energy? Perhaps dozens; we shall never really know.

4) Most importantly, humans have a highly effective intelligent system on which to model their computers: themselves. Natural evolution had no model to work from.

 

It's a far cry from working from scratch.

 

I agree that it appears an appealing approach to use genetic algorithms to create AI, and I can assure you researchers are working in that field.

 

One of my graduate professors worked in that field and he taught several courses where you get to work with genetic algorithms.

 

Not trying to change your mind about concentrating on that field because your right, you may have some unique idea/perspective, but genetic algorithms are definitely researched and studied, and used.

 

But pertaining to my example earlier......

 

Human brains evolved like you say over billions of years, and yet some human brains are not capable of solving or designing only moderately complex algorithms and problems. So maybe the best brains can't solve the AI "algorithm".

 

Lets say we have another million years of evolution, then we will be smart enough to solve it right? We'll not necessarily since at this point in time natural selection in humans is pretty much non-existent.

 

It may require a total wipe, and then maybe still the AI "problem" can't be solved.

 

to grow a human from a fertilized zygote inside a computer.

 

Uh....what?

 

Take a fertilized human zygote, place it inside a "computer" and grow the human there?

 

So you get a computer case with arms, legs, and a head sticking out?

[Luminal]

Well, I do believe my idea is unique because is it highly cross-discipline.

 

I believe research in single disciplines has been performed far too extensively, primarily because of the effects of job specialization in the modern world. For example, the amount of education required to receive a doctorate in both Biochemistry and Computer Science is daunting, to say the least.

 

Thus, modern science has created a massive wealth of information in each respective field, just waiting for people to come along and combine information from multiple fields to produce novel technology.

 

And the crossover that I'm most interested in is bewteen Genetic Algorithms (along with other common Machine Learning algorithms), Neuroscience, and Chemistry. In a sense, the entire brain is a large GA, strengthening certain synaptic connections while weakening others. Yet as I said, I don't want to get overly specific for the remote chance that I have a unique idea.

[bascule]

Generations are arbitrarily short in a computer. Generations in mammals are measured in years, not microseconds.

 

Mammals also possess the most powerful neuroanatomy nature ever conceived of in the form of the neocortex, and more specifically the neocortical column. Mammals possess the most advanced brain structure for building abstract models of the world of any class of life.

 

Uh....what?

 

Take a fertilized human zygote, place it inside a "computer" and grow the human there?

 

So you get a computer case with arms, legs, and a head sticking out?

 

I was thinking more along the lines of a molecular simulation, which would be a hell of a lot more effective if there were a lookup table for how proteins fold (otherwise you need an atomic simulation which can account for Gibbs free energy)

[Fred56]

At present, the map of the cortex (never mind -sorry about that pun- the rest of the brain anatomy) isn't exactly complete.

And we have only rudimentary understanding of how the cortex functions. We still have no good understanding of how any sensory input is encoded, or processed (beyond watching activity using fMRI, or PET), so we certainly need to find out a lot more. And we need the technology that will get us "in" there.

 

This sounds, to me, like 'electronic', or equivalent, devices that will be interfaced directly. This is how we will get whatever is happening in there onto some kind of blackboard, and I also think this is probably inevitable (and it will allow the converse, a brain communicating with a machine directly -like by thinking, something available right now).

 

Current IS models are somewhere along the path, but where exactly? What sort of changes will be brought by the eventual introduction of brain/machine direct interfaces?

 

As long as you aren't too upset about having to plug yourself in every now and then.

[bascule]

At present, the map of the cortex (never mind -sorry about that pun- the rest of the brain anatomy) isn't exactly complete.

 

The specific structure of the neocortical hierarchy is known, however, specifically that sensory and motor areas are on the bottom, the hippocampus is on top, and the interconnecting structures which comprise the hierarchy are neocortical columns, with mesh-like interaction through the thalamus in the form of thalamo (and corticothalamocortical) loops.

 

And we have only rudimentary understanding of how the cortex functions.

 

If you ascribe to what Hawkins' is saying, the same type of structures which wiggle your pinky are the same type of structures which process the painful sensation of stubbing your elbow are the same type of structures which appreciate classical music. Neocortical columns handle high level sense processing, motor control, and perception.

 

These structures pass unrecognized patterns up the neocortical hierarchy and propagate predictions back down it.

 

We still have no good understanding of how any sensory input is encoded

 

What about vision? Scientists have managed to reconstruct scenes using electrical probes implanted in the geniculate nucleus region of a cat's thalamus:

 

http://www.jneurosci.org/cgi/reprint/19/18/8036.pdf

 

This sounds, to me, like 'electronic', or equivalent, devices that will be interfaced directly. This is how we will get whatever is happening in there onto some kind of blackboard, and I also think this is probably inevitable (and it will allow the converse, a brain communicating with a machine directly -like by thinking, something available right now).

 

I have much more hope for passive, non-invasive brain/computer interfaces. They're beginning to reach a point where they can start being commercialized:

 

http://www.gtec.at/products/g.MOBIlab/gMOBIlab.htm

 

Current IS models are somewhere along the path, but where exactly? What sort of changes will be brought by the eventual introduction of brain/machine direct interfaces?

 

If you ascribe to Vernor Vinge's whole Singularity idea, brain/computer interfaces are a road to superhumanity.

[Fred56]

Scientists have managed to reconstruct scenes using electrical probes implanted in the geniculate nucleus region of a cat's thalamus:

K, I'll have a look at this, but my first q is: say the cat isn't given a dose of LSD, now explain how what happened in the cat's subjective brain wasn't something like what a dose of LSD would do?