Found the bit about planes/birds vs humans/AI very insightful:
"But it remains controversial whether further progress in AI will benefit from the study of animal brains. Perhaps we have learned all that we need to from animal brains. Just as airplanes are very different from birds, so one could imagine that an intelligent machine would operate by very different principles from those of a biological organism. We argue that this is unlikely because what we demand from an intelligent machine—what is sometimes misleadingly called “artificial general intelligence”—is not general at all; it is highly constrained to match human capacities so tightly that only a machine structured similarly to a brain can achieve it. An airplane is by some measures vastly superior to a bird: It can fly much faster, at greater altitude, for longer distances, with vastly greater capacity for cargo. But a plane cannot dive into the water to catch a fish, or swoop silently from a tree to catch a mouse. In the same way, modern computers have already by some measures vastly exceeded human computational abilities (e.g. chess), but cannot match humans on the decidedly specialized set of tasks defined as general intelligence."
But while the implementation of an aeroplane is quite different from that of a bird (especially the propulsion side), they don't operate on different principles at all. Birds generate lift and control their flight in the same way aeroplanes do, through aerofoil wings and movable control surfaces.
Likewise, even though the implementation of AGI will almost certainly use logic gates on silicon, the principles behind it will probably be fairly close to those behind biological intelligence (unless, despite all our work on the topic, it turns out that there's an easier way to do AGI than the biological way).
Mm but I think you can save the analogy by switching birds to insects, which use very different aerodynamic principles because fluid dynamics depends greatly on scale. I think it's plausible that something similar could happen for AGI. Perhaps the best way to get AGI depends greatly on whether you have 200 petaflops of exact, sequential operations or a hundred billion neurons that are slow, analog, but massively parallel.
Good point. It seems reasonable that the limited intelligence shown by, say, honeybees works dramatically differently than mammalian intelligence, despite sometimes having surprisingly similar outputs at times.
The constraints put on self-driving cars already show that we clearly expect machines to exceed human performances.
We don't want an AI to show all the bias of humans, we mostly don't care about the emotions, the struggle to remember things, the unreliability of memory, imperfect communication, survival instinct, bodily needs.
Our brain evolved to be able to survive, reproduce, be part of a fitter group. None of these parts of the loss function is useful in most AI problems.
We don't ask a robot to drive a forklift like humans do, we want it to be faster (better, stronger) and more precise. To not tire, to not make mistakes, to not bore, to not wonder about the meaning of life.
The more AI advance and the more we understand the human brain, the clearer it seems (to me at least) that the fact that we are able of rational thought at all is accidental to other specialties that our brain has (we struggle to add small numbers, but we have accelerated circuits to read emotions, pose, injuries).
We may want to learn a bit more about the way our social brain works, but for all the rational things, I suspect the way our brain works is pretty hacky.
I'm not convinced that we actually use "rational thought" very much - isn't there a theory that what we think of as "rational thought" is really a narrative constructed after the fact to explain our own deeper thought processes to ourselves in a very post-hoc fashion.
I suspect this is why "old fashioned" symbolic AI floundered - it was trying to mechanise how we think we think (or rather how we used to think we think) rather than the actual underlying processes.
Wouldn't you expect an AGI that can perform human-like actions also be able to exceed humans in speed, accuracy and recall, simply because it is not built of biological matter with evolutionary constraints on storage capacity, senses, size and longevity?
I agree that we're going to expect the best qualities of a machine, with the best qualities of a human, but the more you push towards a human-like mind the more inherent human-like shortcomings will be introduced, but I envision these to be more in its decision making capacity.
I thought that this analogy was due to Patrick Winston - I remember an AAAI paper that argued this; but I can't find it in his bibliography, so it must have been someone else...
>Genome doesn't have sufficient capacity to specify every connection explicitly .. and can encode about 1GB of information (page 6)
but genome (raw DNA sequence) is only small part of much more complex and dynamic system, which includes regulatory networks, metabolic pathways, RNA interference, cell signalling networks and whatnot [0], surely the number of bits that can be encoded by this is much higher than GB, probably by many orders of magnitude. Genome is just a blueprint of machinery that creates another machinery.
This is not to take away from authors point (and complex systems can be created with simple rules and just a few bits) but innate structure itself can be very complex, I'm wondering if there are any neurophysiological studies showing evidence of a built-in universal grammar, Chomsky's old idea [1]
You cannot use 1GB to encode more than 1GB of information. A 1GB zip file does not encode 1TB of information. Sure, you might have a 1GB zip file that expands to 1TB of data but not information. That is because the 1TB of data has lots of redundancy and inefficiency in how the data is represented that allows the high level of compression.
This is only true for closed systems. The nervous system incorporates a lot of external information through learning. Similar to how you need much less information to define a 1TB storage device. The 1GB figure also neglects epigenetics.
For me this is the most interesting article I've seen on HN this month, thanks for posting this. I also love the subtle yet justified nod to Kant in the title.
I'm currently at work so I only read the abstract but isn't this an obvious alternative to how things are learnt when comparing biological neural networks and artificial neural networks? And don't we call some incidences of this "pre-wiring" instinct?
21 comments
[ 3.2 ms ] story [ 54.3 ms ] threadFound the bit about planes/birds vs humans/AI very insightful:
"But it remains controversial whether further progress in AI will benefit from the study of animal brains. Perhaps we have learned all that we need to from animal brains. Just as airplanes are very different from birds, so one could imagine that an intelligent machine would operate by very different principles from those of a biological organism. We argue that this is unlikely because what we demand from an intelligent machine—what is sometimes misleadingly called “artificial general intelligence”—is not general at all; it is highly constrained to match human capacities so tightly that only a machine structured similarly to a brain can achieve it. An airplane is by some measures vastly superior to a bird: It can fly much faster, at greater altitude, for longer distances, with vastly greater capacity for cargo. But a plane cannot dive into the water to catch a fish, or swoop silently from a tree to catch a mouse. In the same way, modern computers have already by some measures vastly exceeded human computational abilities (e.g. chess), but cannot match humans on the decidedly specialized set of tasks defined as general intelligence."
I hope the article answers these questions because the quote does not...
Also we want to limit AGIs energy usage so we dont have to carry so much power around in mobile versions of it.
Likewise, even though the implementation of AGI will almost certainly use logic gates on silicon, the principles behind it will probably be fairly close to those behind biological intelligence (unless, despite all our work on the topic, it turns out that there's an easier way to do AGI than the biological way).
We don't want an AI to show all the bias of humans, we mostly don't care about the emotions, the struggle to remember things, the unreliability of memory, imperfect communication, survival instinct, bodily needs.
Our brain evolved to be able to survive, reproduce, be part of a fitter group. None of these parts of the loss function is useful in most AI problems.
We don't ask a robot to drive a forklift like humans do, we want it to be faster (better, stronger) and more precise. To not tire, to not make mistakes, to not bore, to not wonder about the meaning of life.
The more AI advance and the more we understand the human brain, the clearer it seems (to me at least) that the fact that we are able of rational thought at all is accidental to other specialties that our brain has (we struggle to add small numbers, but we have accelerated circuits to read emotions, pose, injuries).
We may want to learn a bit more about the way our social brain works, but for all the rational things, I suspect the way our brain works is pretty hacky.
I suspect this is why "old fashioned" symbolic AI floundered - it was trying to mechanise how we think we think (or rather how we used to think we think) rather than the actual underlying processes.
I agree that we're going to expect the best qualities of a machine, with the best qualities of a human, but the more you push towards a human-like mind the more inherent human-like shortcomings will be introduced, but I envision these to be more in its decision making capacity.
but genome (raw DNA sequence) is only small part of much more complex and dynamic system, which includes regulatory networks, metabolic pathways, RNA interference, cell signalling networks and whatnot [0], surely the number of bits that can be encoded by this is much higher than GB, probably by many orders of magnitude. Genome is just a blueprint of machinery that creates another machinery.
This is not to take away from authors point (and complex systems can be created with simple rules and just a few bits) but innate structure itself can be very complex, I'm wondering if there are any neurophysiological studies showing evidence of a built-in universal grammar, Chomsky's old idea [1]
[0] https://en.wikipedia.org/wiki/Systems_biology
[1] https://en.wikipedia.org/wiki/Universal_grammar
In neuroevolution it's called indirect encoding: https://en.wikipedia.org/wiki/Neuroevolution#Direct_and_indi...
"Language Universals Engage Broca's Area" https://journals.plos.org/plosone/article?id=10.1371/journal...
"Language universals at birth" https://www.pnas.org/content/early/2014/03/26/1318261111
"The Effect of Sonority on Word Segmentation" https://onlinelibrary.wiley.com/doi/full/10.1111/j.1551-6709...
"The neurology of syntax" http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.384...
"Cortical tracking of hierarchical linguistic structures": https://www.nature.com/articles/nn.4186
"Mapping syntax using imaging": https://www.ling.upenn.edu/~embick/mappingsyntax.pdf
"Neural Syntax" https://www.researchgate.net/publication/47644138_Neural_Syn...
David Poeppel's lab: https://scholar.google.com/citations?user=9EyT1mYAAAAJ&hl=en...
Yosef Grodzinsky' lab: https://scholar.google.com/citations?user=Qe83EZkAAAAJ&hl=en...