3 ms·

The article looks at an issue I think about quite often and occasionally have a chance to discuss. The complexity of real-world natural systems is at a magnitu

by jrapdx3 10y ago

The article looks at an issue I think about quite often and occasionally have a chance to discuss. The complexity of real-world natural systems is at a magnitude that's hard to grasp and harder to describe. AI, as an attempt to replicate natural human intelligence is up against neural systems that have evolved for 100's of millions of years into numerous forms of astonishingly intricate structure and function. In humans neuronal brain circuitry comprises an astronomical number of interactive elements, and each connection is itself remarkably adaptable with a large number of interfaces associated with a range of signalling modalities. Consider that we have ~100 billion brain neurons and 100-500 trillion synaptic interconnections and each connection is modulated by many factors including endocrine and immune system input. Furthermore connections transmit and receive many kinds of signals in complex feedback relationships. It's well beyond my ability to know how to determine an estimate of how many possible modes of computation can exist within such a marvelous creation, let alone begin to adequately understand how even a small part of it actually works. I suppose that's the author's point, that what we regard as "simple" problems are in reality not at all simple. Computers can do remarkable things, and surely are among the most useful inventions of human beings. We should realize that working like a significant subset of what our brains routinely do isn't going to be easy for computing to achieve.

visarga 10y ago
We don't need to match 1 artificial neuron for 1 biological neuron. An artificial neural network doesn't have to be implemented in a self-replicating body as our brain does. It doesn't have to carry its biological baggage into the computational aspects. Our brain are part of a process of evolution - the specific requirements for that goal impose huge restrictions on it. On the other hand, in a computer, we can have perfect neurons that don't age or degrade, with any structure we want, scalable to a higher degree. We have already seen in image recognition and go play super-human performance, I expect all domains to be conquered soon. Consider go - they started on this project only 2-3 years ago and already AI has overgrown humans. It doesn't need a lot of time to learn once it is created.
- nekopa 10y ago
  I see it as similar to flight, once we stopped trying to mimic nature exactly (flapping planes) we started to not only make progress, but went beyond what nature can do in flight (eg supersonic) I think that the biggest breakthrough we'll have in AI will come from a similar break away from copying the brain. Inspired by nature, not dictated by it.
- jrapdx3 10y ago
  Yes of course you are right that an "artificial neuron" wouldn't (and can't) correspond 1:1 to a biological unit. Main point isn't about aging or degradation because these system characteristics will vary with the nature of the system. However the computational capability of biological neuronal circuits is not very well understood, but it surely is vast, considering the roles subserved for the purposes of the organism in which it is embedded. Personally, like the author of the article, I believe it's naive to predict AI will surpass human brains insofar as the functions that define "human". I don't doubt that computing systems can perform many tasks humans can't easily accomplish, e.g., calculate pi to 1000 places. But I don't regard that as the crux of the question. Determining "meaning" in social interactions, receptive decoding of emotional significance of subtle verbal expression are a couple of examples, but I'll concede they could be unfair. I've been a keen student of human brain structure and function in relation to behavior for a number of decades, and so little is truly known about most all of it. We don't know how any of the several types of memory work, we are just beginning to sort out the mind-boggling complexities of how critical neurons receive and transmit signals. Believe me, I have enormous respect for and great affection for the capabilities of our advancing technologies. I'm a real fan, the reason I very much enjoy HN. But in fundamental ways what we do with computers (and the other relevant technologies) is kind of orthogonal to our neurobiological systems. Conflating these can mislead us to our disadvantage because then what we'd call AI will not really replace "NI" (natural intelligence), imagining it will predisposes to disappointing failures. That's my view of it based on experience, everything has limits, just as you said, including limits on what's humanly knowable, bounded by structural constraints of what we cannot know. Where that boundary lies is indeterminate but not infinite, we keep up the push until slamming into Nature's wall, regroup and turn a different way.
daveguy 10y ago
> Consider that we have ~100 billion brain neurons and 100-500 trillion synaptic interconnections and each connection is modulated by many factors including endocrine and immune system input. Furthermore connections transmit and receive many kinds of signals in complex feedback relationships. This is exactly why estimates of AGI in the next 20 years are laughable. It isn't the neuron that we have to emulate. It is the communication between neurons that we have to emulate. It is absolutely correct that a 1:1 artificial to natural neuron will not be the case. Current artificial neurons (all deep learning neurons) are nowhere near as complex as a natural neuron. It will more likely be 10-100:1 artificial neurons for each neuron. Bandwidth to communicate information between neural processing will be the bottleneck (rather than operations per second). Bits per second is multiple orders of magnitude higher than operations per second in human intelligence and our available bandwidth is at least an order of magnitude lower than our operations per second. That means we are at least 10,000 fold farther in terms of raw processing tech than the most optimistic estimates of ~20 years. 10,000 fold. That is 13-14 doubling times or 26-28 years longer than the most optimistic estimates. That is assuming something like Moore's law holds for technology over the next 46-48 years. That is the time until the fastest supercomputers will be able to emulate one human mind assuming we have figured out the algorithms. 40-60 years is a more reasonable time frame for one human mind on one supercomputer. Everyone should have a pretty good feel for the worth of a prediction that will happen 50 years in the future. There will be at least two more AI winters.

by jrapdx3 10y ago

visarga 10y ago

We don't need to match 1 artificial neuron for 1 biological neuron. An artificial neural network doesn't have to be implemented in a self-replicating body as our brain does. It doesn't have to carry its biological baggage into the computational aspects. Our brain are part of a process of evolution - the specific requirements for that goal impose huge restrictions on it. On the other hand, in a computer, we can have perfect neurons that don't age or degrade, with any structure we want, scalable to a higher degree. We have already seen in image recognition and go play super-human performance, I expect all domains to be conquered soon. Consider go - they started on this project only 2-3 years ago and already AI has overgrown humans. It doesn't need a lot of time to learn once it is created.

nekopa 10y ago
I see it as similar to flight, once we stopped trying to mimic nature exactly (flapping planes) we started to not only make progress, but went beyond what nature can do in flight (eg supersonic) I think that the biggest breakthrough we'll have in AI will come from a similar break away from copying the brain. Inspired by nature, not dictated by it.
jrapdx3 10y ago
Yes of course you are right that an "artificial neuron" wouldn't (and can't) correspond 1:1 to a biological unit. Main point isn't about aging or degradation because these system characteristics will vary with the nature of the system. However the computational capability of biological neuronal circuits is not very well understood, but it surely is vast, considering the roles subserved for the purposes of the organism in which it is embedded. Personally, like the author of the article, I believe it's naive to predict AI will surpass human brains insofar as the functions that define "human". I don't doubt that computing systems can perform many tasks humans can't easily accomplish, e.g., calculate pi to 1000 places. But I don't regard that as the crux of the question. Determining "meaning" in social interactions, receptive decoding of emotional significance of subtle verbal expression are a couple of examples, but I'll concede they could be unfair. I've been a keen student of human brain structure and function in relation to behavior for a number of decades, and so little is truly known about most all of it. We don't know how any of the several types of memory work, we are just beginning to sort out the mind-boggling complexities of how critical neurons receive and transmit signals. Believe me, I have enormous respect for and great affection for the capabilities of our advancing technologies. I'm a real fan, the reason I very much enjoy HN. But in fundamental ways what we do with computers (and the other relevant technologies) is kind of orthogonal to our neurobiological systems. Conflating these can mislead us to our disadvantage because then what we'd call AI will not really replace "NI" (natural intelligence), imagining it will predisposes to disappointing failures. That's my view of it based on experience, everything has limits, just as you said, including limits on what's humanly knowable, bounded by structural constraints of what we cannot know. Where that boundary lies is indeterminate but not infinite, we keep up the push until slamming into Nature's wall, regroup and turn a different way.

daveguy 10y ago

> Consider that we have ~100 billion brain neurons and 100-500 trillion synaptic interconnections and each connection is modulated by many factors including endocrine and immune system input. Furthermore connections transmit and receive many kinds of signals in complex feedback relationships. This is exactly why estimates of AGI in the next 20 years are laughable. It isn't the neuron that we have to emulate. It is the communication between neurons that we have to emulate. It is absolutely correct that a 1:1 artificial to natural neuron will not be the case. Current artificial neurons (all deep learning neurons) are nowhere near as complex as a natural neuron. It will more likely be 10-100:1 artificial neurons for each neuron. Bandwidth to communicate information between neural processing will be the bottleneck (rather than operations per second). Bits per second is multiple orders of magnitude higher than operations per second in human intelligence and our available bandwidth is at least an order of magnitude lower than our operations per second. That means we are at least 10,000 fold farther in terms of raw processing tech than the most optimistic estimates of ~20 years. 10,000 fold. That is 13-14 doubling times or 26-28 years longer than the most optimistic estimates. That is assuming something like Moore's law holds for technology over the next 46-48 years. That is the time until the fastest supercomputers will be able to emulate one human mind assuming we have figured out the algorithms. 40-60 years is a more reasonable time frame for one human mind on one supercomputer. Everyone should have a pretty good feel for the worth of a prediction that will happen 50 years in the future. There will be at least two more AI winters.