I can see the utility of this from the point of view of trigonometry. At one point in the chapter he gives an example of a simple problem that is needlessly complex to do using tan (if you dont have a calculator.) As such a knowledge of this would surely be useful.
However it is not clear to me how you could use spread to map a continuiously increasing quantity (such as time) on to a periodic variable (such as displacement.) Surely to do this using his simple ratio of quadrances would be more complex than using sin? Then what about things like Fourier series? This would surely be very clunky in this framework.
This stuff must still be equivalent to classical trig. Thus it cant possibly be 'revolutionary'. You still need to start with the same axioms.
Seems to me that the big players like M$ and $ony have already got this figured. The XBox 360 and PS3 might become the first actually successful 'net PCs' when paired with broadband connections. And they will be around the magic price-point too.
The new consoles will be computing appliances because people will be able to have their media, their games and their email/web access in one place on hardware that will be pretty constrained and hence easy to use out of the box. No doubt they will also be practically 'instant-on' like Macs can be.
Infact as we all know Sony are keen to get the PS3 seen as a 'real' computer.
PCs as we know them will be left for professionals, businesses and people who actually *enjoy* administering their systems.
I agree with you that you cannot easily isolate behavoir from learning. And it is certianly the case that the Spike timing plasticity results are a little contrived.
I guess this thing wont be conscious after all then!
"There are many examples of important connections between neurons with cell bodies spaced millimeters apart. Like the retinal to LGN connections without which you cannot see."
Yes but they are not non-local in the neural network sense of the word are they? And that is what we were taking about (*learning-rules*). Non-local in terms of learning rules means such pracitises as wieght renormalisation which are non-physical and non-biological. Non-local is not the same thing as long range or 'top-down' which is the issue you are addressing with your reference to the LGN so it seems you have your concepts muddled.
"There are senior neuroscientists working with senior machine learning experts to create machine learning algorithms that emulate the brain. Right now. You just didn't read about it on/."
I think you just blinded me with science. What machine learning algorithms attempt to 'emulate the brain'? No one I know of in machine learning does *anything* with the explicity intention of 'emulating the brain'. They might do things like perceptual inference but that is something different again. Besides such endevours use probabilistic models and graphical models which are not the same thing as neural networks and learning rules so once again you are muddled.
"Non-locality doesn't mean go back to physics. It means that there are really long range connections that are really important for learning in the brain, and you cannot study those in a slice."
If you mean non-locality to be synominous with long-range then I agree. But since the orignal context was to do with ANN learning rules in which non-locailty means the mathematical property of certain weights effecting others with no explicit causuality built in to the model, you are using the term incorrectly. Also I dont see why you cant study long-range connections in slices; you can. you just cant do it a way that is demonstrably physiological which is probably the point you were making.
"STDP is not so difficult it cannot be studied. It just doesn't work the way the slice folks think it does when you are in a living organism and not in a slice."
Ah yes, that was the point you were trying to make. Well I am afraid that this point applies to most of neuroscience. Again it is something we have live with.
What do you mean 'out of his league talking about AI'. He isnt talking about AI, he is talking about neuroscience!
Besides, the idea is that there is going to be *covergence* between AI and neuroscience in as much as you ever get convergence between pure sciences like Neuroscience or Computational-Neuroscience and what are essentially engineering practices such as 'AI'
Dude I dont know what you are on about: "There are SOME local learning rules, but there are important neural networks learning rules that are distinctly non-local."
Are you now advocating non-locality? Thats getting beyond neuroscience and in to physics! It really is very likely that the rules will have to be local. Life might immitate art but there is no reason whatsoever that the brain should imitate 'AI' or even neural networks for that matter. Neural networks are polynomials. Brains are not. Non-locality might exist in a computer but it seems unlikely that it exists in the universe at least not on the level of a large hot macroscopic lump of matter like a brain (unless you believe Penrose of course, but I dont)
The fact that STDP has not been confirmed in vivo is not an arguement for the simplicity of neurons now is it? There are myriad things that have not been confirmed in vivo but this largely to the technical difficulty of performing such experiments. We just have to live with that.
Hold your horses! There is abundant evidence that single neurons can perform more complex operations than a mere 'sigmoid fuunction'. That is a working approximation that can be useful from the point of view of simulations but that is all.
Single neurons can potentially perform computations at the level of the of the passive cable equation. At the level of active membrane properties when added to those passive canle equation solutions. At the level of genetic instructions becoming activated in the nucleus and dendrites in response to activity. And finally the plasticity or learning rules that neurons use are not only computational very important but probably quite varied from brain region to region. Spike timing dependent plasticity for example allows the brain to pick out persistent correlations within highly noisy inputs. None of this is included in the impoverised neural-network viewpoint of 'sigmoids'
The real question is why are they doing this? Markram is a top researcher and knows what he is doing. But i quesiton the motivations of big blue. i wouldnt be suprised if they didnt give two hoots about the science but rather are only doing this so that they can get the kind of publicity that posts on slashdot bring. Remember 'Deep Blue'? Lets hope they dont treat Markram like they did Kasparov
Yes but if they had never privatised BT I doubt we would even *have* broadband. We would probably have resorted to binary semaphore, or perhaps smoke signals on a still day.
Cos the socialist state that preceeded Thatcher was perfect wasn't it?
(Also, personally speaking, my broadband expereinces with BT have been perfectly adequate)
With the new breed of VGA PDAs, do we not already have THGTTG in our pockets?
Surfing in true VGA mode on the 4700 means you can look at things like wikipedia on the move anyway as long as you have access to a local wireless LAN and/or a bluetooth phone.
There are ways in which PDAs and the 4700 suck. However in my experience one thing that VGA on the 4700 *does* offer is genuine desktop-like anywhere internet reference.
Slashdot Mirror
I can see the utility of this from the point of view of trigonometry. At one point in the chapter he gives an example of a simple problem that is needlessly complex to do using tan (if you dont have a calculator.) As such a knowledge of this would surely be useful.
However it is not clear to me how you could use spread to map a continuiously increasing quantity (such as time) on to a periodic variable (such as displacement.) Surely to do this using his simple ratio of quadrances would be more complex than using sin? Then what about things like Fourier series? This would surely be very clunky in this framework.
This stuff must still be equivalent to classical trig. Thus it cant possibly be 'revolutionary'. You still need to start with the same axioms.
Nobody has mentioned the PS3 or Xbox yet...
Seems to me that the big players like M$ and $ony have already got this figured. The XBox 360 and PS3 might become the first actually successful 'net PCs' when paired with broadband connections. And they will be around the magic price-point too.
The new consoles will be computing appliances because people will be able to have their media, their games and their email/web access in one place on hardware that will be pretty constrained and hence easy to use out of the box. No doubt they will also be practically 'instant-on' like Macs can be.
Infact as we all know Sony are keen to get the PS3 seen as a 'real' computer.
PCs as we know them will be left for professionals, businesses and people who actually *enjoy* administering their systems.
I agree with you that you cannot easily isolate behavoir from learning. And it is certianly the case that the Spike timing plasticity results are a little contrived.
I guess this thing wont be conscious after all then!
"There are many examples of important connections between neurons with cell bodies spaced millimeters apart. Like the retinal to LGN connections without which you cannot see."
Yes but they are not non-local in the neural network sense of the word are they? And that is what we were taking about (*learning-rules*). Non-local in terms of learning rules means such pracitises as wieght renormalisation which are non-physical and non-biological. Non-local is not the same thing as long range or 'top-down' which is the issue you are addressing with your reference to the LGN so it seems you have your concepts muddled.
"There are senior neuroscientists working with senior machine learning experts to create machine learning algorithms that emulate the brain. Right now. You just didn't read about it on /."
I think you just blinded me with science. What machine learning algorithms attempt to 'emulate the brain'? No one I know of in machine learning does *anything* with the explicity intention of 'emulating the brain'. They might do things like perceptual inference but that is something different again. Besides such endevours use probabilistic models and graphical models which are not the same thing as neural networks and learning rules so once again you are muddled.
"Non-locality doesn't mean go back to physics. It means that there are really long range connections that are really important for learning in the brain, and you cannot study those in a slice."
If you mean non-locality to be synominous with long-range then I agree. But since the orignal context was to do with ANN learning rules in which non-locailty means the mathematical property of certain weights effecting others with no explicit causuality built in to the model, you are using the term incorrectly. Also I dont see why you cant study long-range connections in slices; you can. you just cant do it a way that is demonstrably physiological which is probably the point you were making.
"STDP is not so difficult it cannot be studied. It just doesn't work the way the slice folks think it does when you are in a living organism and not in a slice."
Ah yes, that was the point you were trying to make. Well I am afraid that this point applies to most of neuroscience. Again it is something we have live with.
What do you mean 'out of his league talking about AI'. He isnt talking about AI, he is talking about neuroscience!
Besides, the idea is that there is going to be *covergence* between AI and neuroscience in as much as you ever get convergence between pure sciences like Neuroscience or Computational-Neuroscience and what are essentially engineering practices such as 'AI'
Dude I dont know what you are on about: "There are SOME local learning rules, but there are important neural networks learning rules that are distinctly non-local."
Are you now advocating non-locality? Thats getting beyond neuroscience and in to physics! It really is very likely that the rules will have to be local. Life might immitate art but there is no reason whatsoever that the brain should imitate 'AI' or even neural networks for that matter. Neural networks are polynomials. Brains are not. Non-locality might exist in a computer but it seems unlikely that it exists in the universe at least not on the level of a large hot macroscopic lump of matter like a brain (unless you believe Penrose of course, but I dont)
The fact that STDP has not been confirmed in vivo is not an arguement for the simplicity of neurons now is it? There are myriad things that have not been confirmed in vivo but this largely to the technical difficulty of performing such experiments. We just have to live with that.
Hold your horses! There is abundant evidence that single neurons can perform more complex operations than a mere 'sigmoid fuunction'. That is a working approximation that can be useful from the point of view of simulations but that is all.
Single neurons can potentially perform computations at the level of the of the passive cable equation. At the level of active membrane properties when added to those passive canle equation solutions. At the level of genetic instructions becoming activated in the nucleus and dendrites in response to activity. And finally the plasticity or learning rules that neurons use are not only computational very important but probably quite varied from brain region to region. Spike timing dependent plasticity for example allows the brain to pick out persistent correlations within highly noisy inputs. None of this is included in the impoverised neural-network viewpoint of 'sigmoids'
The real question is why are they doing this? Markram is a top researcher and knows what he is doing. But i quesiton the motivations of big blue. i wouldnt be suprised if they didnt give two hoots about the science but rather are only doing this so that they can get the kind of publicity that posts on slashdot bring. Remember 'Deep Blue'? Lets hope they dont treat Markram like they did Kasparov
Yes but if they had never privatised BT I doubt we would even *have* broadband. We would probably have resorted to binary semaphore, or perhaps smoke signals on a still day.
Cos the socialist state that preceeded Thatcher was perfect wasn't it?
(Also, personally speaking, my broadband expereinces with BT have been perfectly adequate)
With the new breed of VGA PDAs, do we not already have THGTTG in our pockets?
Surfing in true VGA mode on the 4700 means you can look at things like wikipedia on the move anyway as long as you have access to a local wireless LAN and/or a bluetooth phone.
There are ways in which PDAs and the 4700 suck. However in my experience one thing that VGA on the 4700 *does* offer is genuine desktop-like anywhere internet reference.