I have no doubt in my mind that some time in the future nervous systems with be simulated with all their functions including consciousness. Perhaps not a particular person’s nervous system at a particular time, but a somewhat close approximation, a very similar nervous system with consciousness but no magic. However, I definitely doubt that it will be done on a general purpose computer running algorithms. I doubt that step-by-step calculations will be the way that the simulation will be done. Here is why:
1.The brain is massively parallel and complex feedback loops are difficult to calculate (not impossible but difficult). The easiest way to simulate a massively parallel system is to build it in hardware rather than use stepwise software.
2.There are effects of fields to consider – not just electrical and magnetic but also chemical. Like massive numbers of feedback loops, the fields would be difficult to calculate as the same elements that are reacting to the fields are also creating them.
3.There are many critical timing effects in the system and these would have to be duplicated or scaled, another difficulty of calculation.
I believe that it is far simpler to take advantage of the architecture of the brain which appears to have a lot of repetition of small units of a few thousand cells and build good models of these in hardware, including correct timing and ways to simulate fields etc. Then take advantage of the larger (sort of functional) divisions of the brain to construct larger modules. It gets very complicated fairly quickly but not as complicated as stepwise calculations. In essence it resembles the replacement of neurons one at a time with chips but the chips would have to be more more than just fancy logic components as they would have to sense their surroundings as well as communicate with other neurons or chips. The boundaries need to be at the natural joints to make it simpler, but the idea is the same. I can imagine this actually being built and having consciousness. The computer running algorithms or the person with a pencil creating consciousness is a lot harder to imagine (and needs a lot of ‘in principles’, too many for me).
Hardware might ultimately be more efficient than software for this kind of thing, but software is a lot easier to tune and debug. There are reasons neural network chips never took off.
I can plausibly imagine the first upload running in software, orders of magnitude slower than real time, on enough computers to cover a city block and require a dedicated power station, cooperating with a team of engineers and neuroscientists by answering one test question per day; 10 years later, the debugged version implemented in hardware, requiring only a roomful of equipment per upload, and running at a substantial fraction of real-time speed; and another 10 years later, new process technology specifically designed for that hardware, allowing a mass-market version that runs at full real-time speed, fits in desktop form factor and plugs into a standard power socket.
You may be right but my imagination has a problem with it. If there is a way to do analog computing using software in a non step-by-step procedure, then I could imagine a software solution. It is the algorithm that is my problem and not the physical form of the ‘ware’.
I may not be understanding your objection in that case. Are you saying that there’s no way software, being a digital phenomenon, can simulate continuous analog phenomena? If so, I will point to the many cases where we successfully use software to simulate analog phenomena to sufficient precision. If not, can you perhaps rephrase?
I may not be expressing my self well here. I am try to express what I can and cannot imagine—I do not presume to say that because I cannot imagine something, it is impossible. In fact I believe that it would be possible to simulate the nervous system with digital algorithms in principle, just extremely difficult in practice. So difficult I think that I cannot imagine it happening. It is not the ‘software’ or the ‘digital’ that is my block, it is the ‘algorithm’, the stepwise processes that I am having trouble with. How do you imagine the enormous amount and varied nature of feedback in the brain can be simulated by step-by-step logic? I take it that you can imagine how it could be done—so how?
I guess that is the conversation stopper. We agree that it takes a lot of steps. We disagree on whether the number makes it only possible in principle or not.
Ah, I was about to reply with a proof of concept explanation in terms of molecular modeling (which of course would be hopelessly intractable in practice but should illustrate the principle), until I saw you say ‘only possible in principle’; are you saying then that your objection is that you think even the most efficient software-based techniques would take, say, a million years of supercomputer time to run a few seconds of consciousness?
Well, maybe not that long, but a long, long time to do the ‘lot of little steps’. It does not seem the appropriate tool to me. After all, the much slower component parts of a brain do a sort of unit of perception in about a third of a second. I believe that is because it is not done step-wise but something like this: the enormous number of overlapping feedback loops can only stabilize in a sort of ‘best fit scenario’ and it takes very little time for the whole network to hone in on the final perception. (Vaguely that sort of thing)
Right, fair enough, then it’s a quantitative question on which our intuitions differ, and the answer depends both on a lot of specific facts about the brain, and on what sort of progress Moore’s Law ends up making over the next few decades. Let’s give Blue Brain another decade or two and see what things look like then.
Personally I have great hopes for Blue Brain. If it figures out how a single cortex unit works ( which they seem to be on the way to). If they can then figure out how to convert that into a chip and put oodles of those clips in the right environment of inputs and interactions with other parts of the brain (thalamus and basal ganglia especially) and then.....
A lot of work but it has a good chance as long as it avoids the step-by-step algorithm trap.
I have no doubt in my mind that some time in the future nervous systems with be simulated with all their functions including consciousness. Perhaps not a particular person’s nervous system at a particular time, but a somewhat close approximation, a very similar nervous system with consciousness but no magic. However, I definitely doubt that it will be done on a general purpose computer running algorithms. I doubt that step-by-step calculations will be the way that the simulation will be done. Here is why:
1.The brain is massively parallel and complex feedback loops are difficult to calculate (not impossible but difficult). The easiest way to simulate a massively parallel system is to build it in hardware rather than use stepwise software.
2.There are effects of fields to consider – not just electrical and magnetic but also chemical. Like massive numbers of feedback loops, the fields would be difficult to calculate as the same elements that are reacting to the fields are also creating them.
3.There are many critical timing effects in the system and these would have to be duplicated or scaled, another difficulty of calculation.
I believe that it is far simpler to take advantage of the architecture of the brain which appears to have a lot of repetition of small units of a few thousand cells and build good models of these in hardware, including correct timing and ways to simulate fields etc. Then take advantage of the larger (sort of functional) divisions of the brain to construct larger modules. It gets very complicated fairly quickly but not as complicated as stepwise calculations. In essence it resembles the replacement of neurons one at a time with chips but the chips would have to be more more than just fancy logic components as they would have to sense their surroundings as well as communicate with other neurons or chips. The boundaries need to be at the natural joints to make it simpler, but the idea is the same. I can imagine this actually being built and having consciousness. The computer running algorithms or the person with a pencil creating consciousness is a lot harder to imagine (and needs a lot of ‘in principles’, too many for me).
Hardware might ultimately be more efficient than software for this kind of thing, but software is a lot easier to tune and debug. There are reasons neural network chips never took off.
I can plausibly imagine the first upload running in software, orders of magnitude slower than real time, on enough computers to cover a city block and require a dedicated power station, cooperating with a team of engineers and neuroscientists by answering one test question per day; 10 years later, the debugged version implemented in hardware, requiring only a roomful of equipment per upload, and running at a substantial fraction of real-time speed; and another 10 years later, new process technology specifically designed for that hardware, allowing a mass-market version that runs at full real-time speed, fits in desktop form factor and plugs into a standard power socket.
You may be right but my imagination has a problem with it. If there is a way to do analog computing using software in a non step-by-step procedure, then I could imagine a software solution. It is the algorithm that is my problem and not the physical form of the ‘ware’.
I may not be understanding your objection in that case. Are you saying that there’s no way software, being a digital phenomenon, can simulate continuous analog phenomena? If so, I will point to the many cases where we successfully use software to simulate analog phenomena to sufficient precision. If not, can you perhaps rephrase?
I may not be expressing my self well here. I am try to express what I can and cannot imagine—I do not presume to say that because I cannot imagine something, it is impossible. In fact I believe that it would be possible to simulate the nervous system with digital algorithms in principle, just extremely difficult in practice. So difficult I think that I cannot imagine it happening. It is not the ‘software’ or the ‘digital’ that is my block, it is the ‘algorithm’, the stepwise processes that I am having trouble with. How do you imagine the enormous amount and varied nature of feedback in the brain can be simulated by step-by-step logic? I take it that you can imagine how it could be done—so how?
with a lot of steps.
I guess that is the conversation stopper. We agree that it takes a lot of steps. We disagree on whether the number makes it only possible in principle or not.
Ah, I was about to reply with a proof of concept explanation in terms of molecular modeling (which of course would be hopelessly intractable in practice but should illustrate the principle), until I saw you say ‘only possible in principle’; are you saying then that your objection is that you think even the most efficient software-based techniques would take, say, a million years of supercomputer time to run a few seconds of consciousness?
Well, maybe not that long, but a long, long time to do the ‘lot of little steps’. It does not seem the appropriate tool to me. After all, the much slower component parts of a brain do a sort of unit of perception in about a third of a second. I believe that is because it is not done step-wise but something like this: the enormous number of overlapping feedback loops can only stabilize in a sort of ‘best fit scenario’ and it takes very little time for the whole network to hone in on the final perception. (Vaguely that sort of thing)
Right, fair enough, then it’s a quantitative question on which our intuitions differ, and the answer depends both on a lot of specific facts about the brain, and on what sort of progress Moore’s Law ends up making over the next few decades. Let’s give Blue Brain another decade or two and see what things look like then.
Personally I have great hopes for Blue Brain. If it figures out how a single cortex unit works ( which they seem to be on the way to). If they can then figure out how to convert that into a chip and put oodles of those clips in the right environment of inputs and interactions with other parts of the brain (thalamus and basal ganglia especially) and then.....
A lot of work but it has a good chance as long as it avoids the step-by-step algorithm trap.