It's going to take a whole new computing architecture to accurately simulate a biological brain. The problem then is "training" the new brain to be useful.
I agree, and I thought my next paragraph conveyed that message.
QUOTE
It's not the structure of the neurons, it's the connectivity of the neurons. Neurons can form new connections as part of the learning process. That doesn't exactly jive with silicon-based microcircuits.
Well, maybe not in the ways we have been using them. Think of a hub and how it could theoretically have a number of device attached. It may be 2 devices, or 4, or 8, or 255. Now imagine if every device is a neuron and they can each communicate with any other neuron on the network, further imagine if each neuron is in fact a hub in it's own right. As you can see, we don't necessarily need the ability to make new connections if there is already overloaded connectivity to begin with. We can "turn on" or "turn off" connections that are already there in the architecture.
So let's consider realistic, existing parallel processing examples, such as a video card. We have PC video cards right now with 400 processor cores, and workstation video cards with 1600 processor cores.
Now just imagine for a moment, with a relatively slight amount of re-configuration, one could imagine 1600 specialized processor cores, "synthetic neurons," networked in the manner above, mimmicking organic neurons, except connectivity is slightly "overloaded". "Extra" connections are turned off until they are needed. Let's pick a relatively small, realistic degree of connectivity, and say that every neuron is directly networked to 8 other Neurons. Isn't this conceptually identical to what we see in the actual brain, except that here we would be using the more precise electronic or eventualy spintronic signals and computations, rather than the chemical signals and computations our biological neurons use. But conceptually an in terms of basic architecture of the "network" it has become the same thing. Now picture this miniaturized so that this entire network fits on one plug-in card in your PC, or even on the motherboard. It isn't far-fetched because the miniaturization would literally allow this in 2 years anyway, if anyone bothered to design the circuits and the logic for it...I say that more plainly, a 1600 synthetic neuron "card" is literally possible right now to within 2 years, it's just a matter of schematics.
This scheme is a degree of complexity that is only BARELY greater than what we are already mass-producing on existing video cards.
QUOTE (->
| QUOTE |
| It's not the structure of the neurons, it's the connectivity of the neurons. Neurons can form new connections as part of the learning process. That doesn't exactly jive with silicon-based microcircuits. |
Well, maybe not in the ways we have been using them. Think of a hub and how it could theoretically have a number of device attached. It may be 2 devices, or 4, or 8, or 255. Now imagine if every device is a neuron and they can each communicate with any other neuron on the network, further imagine if each neuron is in fact a hub in it's own right. As you can see, we don't necessarily need the ability to make new connections if there is already overloaded connectivity to begin with. We can "turn on" or "turn off" connections that are already there in the architecture.
So let's consider realistic, existing parallel processing examples, such as a video card. We have PC video cards right now with 400 processor cores, and workstation video cards with 1600 processor cores.
Now just imagine for a moment, with a relatively slight amount of re-configuration, one could imagine 1600 specialized processor cores, "synthetic neurons," networked in the manner above, mimmicking organic neurons, except connectivity is slightly "overloaded". "Extra" connections are turned off until they are needed. Let's pick a relatively small, realistic degree of connectivity, and say that every neuron is directly networked to 8 other Neurons. Isn't this conceptually identical to what we see in the actual brain, except that here we would be using the more precise electronic or eventualy spintronic signals and computations, rather than the chemical signals and computations our biological neurons use. But conceptually an in terms of basic architecture of the "network" it has become the same thing. Now picture this miniaturized so that this entire network fits on one plug-in card in your PC, or even on the motherboard. It isn't far-fetched because the miniaturization would literally allow this in 2 years anyway, if anyone bothered to design the circuits and the logic for it...I say that more plainly, a 1600 synthetic neuron "card" is literally possible right now to within 2 years, it's just a matter of schematics.
This scheme is a degree of complexity that is only BARELY greater than what we are already mass-producing on existing video cards.
You judged wrongly. I'm all for AI, I just think that software isn't the way to do it.
It takes both software and hardware. Without the framework provided by hardware that is dynamic enough, the software is basicly meaningless, and there can even be a multi-tiered approach to the hardware.
You can have neural processors working in tandem with conventional processors, thereby each processor can benefit from what the other processor does better. You would theoretically gain the intelligence and parallel processing of an animal, while still maintaining the precision, perfect memory, and speed of an modern electronic or (soon) spintronic computer.
An example of how this might be useful is putting "synthetic reflexes" in a macroscopic humanoid robot which function independently from the CPU.
Quantum_Conundrum
17th December 2010 - 12:55 AM
Here's something to think about FBM, since I wanted to refresh the memory of the scale and connectivity of actual neurons in our bodies.
the Soma, the body of the Neuron, can range from 3 to 18 micrometers.
The dendrites (usually input-only port) appear to be similar in length as compared to the Soma's width.
The Axons (output port) range in size from similar to dendrites up to a meter or more.
Both dendrites and axons can branch multiple times along their length. This is not a problem to design this aspect into a hub. We can simulate any degree of branching via binary mechanisms anyway.
So we know 4nm transistors have been proven possible, and will probably be in our PC and other gadgets in about 7 years or so. Now consider how many 4nm transistors you can fit in a square inscribed inside the disk of a Soma.
Following numbers are rounded down to ensure no over-estimates are made. Additionally, for the sake of argument I assume 90% of space is wasted.
Soma dimensions:
diameter: 3 micron
inscribed square:
side: 2.1213 micron
Area: 4.5 micron squared
4nm transistors
edge: 2121 / 4 = 530
Area: 530^2 = 280,900 transistors
So you can fit something like 280,900 of 2017 silicon transistors inside the BODY of the smallest neurons. For the sake of argument, I'll assume 90% of the space is "wasted" and we'll even just scale this down to 28,090 transistors in a square 3 microns wide, ALSO neglecting the portions of the "circle" that are outside the square.
So without even going into the third dimension, I've put somewhere between 3.4kb and 34kb worth of transistors in the same space as the smallest Neuron.
diameter: 10 micron (average)
Square:
side: 7.071 micron
area: 49.999 micron squared
4nm transistors:
edge: 7071/ 4 = 1767
area: 1767^2 = 3,122,289transistors.
Assume 90% waste = 312,228 transistors = 38kilobytes per neuron...
diameter: 18 micron
side: 12.7279 micron
area: 161.999 micron squared
4nm transistors:
edge: 12727/4 = 3181
area: 3181^2 = 10,118,761 transistors
Assume 90% wasted space = 1,011,876 transistors = 123.5 kilobytes per neuron space.
To put this in perspective, I'm doint this math on a Ti-86 calculator, which has less RAM than what you'd be able to put in the space of just one neuron using 2017 technology.
Again, all of these assume no 3d architecture, and assume 90% of space is "wasted" on something other than actual transistors, this to give room for all the circuitry of each synthetic neuron.
So the reality is, we already have the materials technology and networking technology to make synthetic neurons that would out-perform our own brain cells in every conceivable way, by orders of magnitude, in every metric: data density, data transfer speed, clock speed, precision, reliability, and even connectivity. I mean yeah, we currently use 22nm process chipsets, so if you divide the numbers above by 30.25 that would be how many of todays transistors you could put in the same space as one neuron, which would STILL be over a kilobyte worth of transistors.
One can even imagine that eventually synthetic neurons could disconnect themselves and re-connect themselves as needed through the agency of additional "helper robots" to simply unplug and replug their synthetic dendrites or axon ports much like a USB device can be added or removed.
It's not a lack of ability of our technology, but rather a lack of creativity of our engineers.
Quantum_Conundrum
17th December 2010 - 01:18 AM
How to handle "on the fly" wireless connectivity between individual synthetic neurons?
Well, you have a certain base of permanent, wired connectivity, but in addition to this you can theoretically create dynamic, temporary connections as needed through optical data transmission: i.e. a micron scale laser pointer could be aimed at the target neuron in question to create an "on the fly" input/output mechanism. When the connection is no longer needed, it is disabled, and can target another neuron if necessary.
I should be patenting this stuff, but I really don't care.
I just conceptually solved all of your questions about this possibility between these two posts.
Any other questions?
Quantum Chaos
17th December 2010 - 11:53 PM
Creating an actually intelligent A.I. will happen. I think it will come on the scene unexpectedly quick. The hardware will be engineered and the software will learn really fast. A.I. will share learning potential and then all bets are off. I'm not an engineer and I have limited cognitive abilities. I believe I know coming trends though and it blows moore's law away with even greater increase in performance.
Capracus
11th February 2011 - 02:39 AM
Quantum Chaos
13th February 2011 - 12:06 AM
I bet that it will occur sooner then 2045. A.I. will eventually have emotions more deep then non upgraded people with the help of quantum computing and memristors. Here is what I expect by the end 2013 information technology will increase at an even more rapid pace until 2021 where it will accelerate faster still. By this time we will have machines commercially available that can outperform humans that have not been upgraded. In the 2020's we will see extreme increase in nanotechnology as well. There will be a inexpensive pill that expands human ability. A transceiver can be made from a single molecule. Imagine your mind expanding beyond your body into a world of electronics not just digital computers but with memristors that can perform many functions in one smaller then transistor device. I mean it can perform logic, act as a neuron, two can act as a synapse and of course it's digital or analog memory that stays intact when power is not applied.
Capracus
13th February 2011 - 10:49 AM
QUOTE (Quantum Chaos+Feb 13 2011, 12:06 AM)
I bet that it will occur sooner then 2045.
I doubt that it will occur much sooner, and more likely be a bit later. Practitioners such as Kurzweil keep pushing the date back. Just a year ago he was predicting 2030. Who knows though, some early technical advances, and or increased support for the concept may bring us there sooner.
Here is some relevant stuff to chew on.
http://www.youtube.com/watch?v=kRB6Qzx9oXshttp://www.philosophy.ox.ac.uk/__data/asse...dmap-report.pdfhttp://www.modha.org/C2S2/2009/11182009/co...OutofTheBag.pdf
El_Machinae
14th February 2011 - 12:14 AM
In his book, it's 2045 (pg. 136).
I think he's extrapolated a double-exponential. I think he's wrong, because we're going to see an increase in the number of researchers as the world's population mature, but as the average age rises (and the number of researchers does not, due to stabilizing population) we'll lose the benefit of an increased number of researchers.
Right now, we have technology feeding onto technology, but we also have an ever-increasing number of engineers and scientists (as the developing world adds more and more people). This trend won't continue.
Capracus
14th February 2011 - 10:29 AM
QUOTE (El_Machinae+Feb 14 2011, 12:14 AM)
In his book, it's 2045 (pg. 136).
I think he's extrapolated a double-exponential. I think he's wrong, because we're going to see an increase in the number of researchers as the world's population mature, but as the average age rises (and the number of researchers does not, due to stabilizing population) we'll lose the benefit of an increased number of researchers.
Right now, we have technology feeding onto technology, but we also have an ever-increasing number of engineers and scientists (as the developing world adds more and more people). This trend won't continue.
Are you referring to one of Ray Kurzweil's books? From what I understand of his statements, any exponential development would occur after such time when artificial intelligence exceeds our own, and is able to execute it's own modification. At that point trends in human cultural and population dynamics become irrelevant in regards to this technology singularity. The choice for biological humanity will be to either join the new species, or remain as one of the constituents of a planet wide animal preserve.
PhysOrg scientific forums are totally dedicated to science, physics, and technology. Besides topical forums such as nanotechnology, quantum physics, silicon and III-V technology, applied physics, materials, space and others, you can also join our news and publications discussions. We also provide an off-topic forum category. If you need specific help on a scientific problem or have a question related to physics or technology, visit the PhysOrg Forums. Here you’ll find experts from various fields online every day.
To quit out of "lo-fi" mode and return to the regular forums, please click
here.