8 ms·

Background: I did AI and Philosophy of Mind in undergrad, an MSc focused on ALife, then a PhD at Yale under a Macarthur Fellow who developed the theoretical fra

Background: I did AI and Philosophy of Mind in undergrad, an MSc focused on ALife, then a PhD at Yale under a Macarthur Fellow who developed the theoretical framework for the 'evolution of Evolvability' where I worked on computational evolutionary biology. I can say we're not going to blindly brute force our way forward, but instead we'll need to reverse engineer nature's core algorithms to generate hard AI. Every time a major advance is made in AI the computational neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!". Those ingredients will be embodiment, evolution, genetics (genotype-phenotype encoding), neirogenesis(gene regulatory networks directing phenotypic development from a single cell to a multicellular neural network), and ecology (evolving in adversarial and cooperative environments). And we'll need a lot of theoretical work in how to represent nature's algorithms in code. For example my PhD work just focused on how to use evolutionary algorithms to evolve simple gene regulatory networks and how that leads to properties of modularity in the genotype-phenotype map. That alone is life's work but a necessary ingredient. I don't expect to see this solved in my lifetime given how we're attacking the problem (head on) today. And until then we're going to continue to run into these dark winters of AI.

heimatau 9y ago
Hmm, reading your experience clarifies to me (a Maths guy) why current deep learning efforts are producing real fruit from their labor. Linear Regression and Vector/dot products are just constructs to re-create the evolutionary biology seen in nature. When looking at the slow process of evolution, it's the nuances that matter and over time, they compound exponentially. Pretty neat. By breaking down a task into abstract constructs and sifting problems out with a sigmoid function.
eggie 9y ago
I'd love to read your thesis. Would you please a link?
Jedi72 9y ago
Can you link to some of your work, this is very interesting stuff
return0 9y ago
> neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!". Interested to hear about these. I can honestly not remember any.
- ghaff 9y ago
  There's been this tension over whether achieving apparent human levels of artificial intelligence requires thinking like and following processes associated with human learning and intelligence pretty much forever. Today, computational capability and data have gotten us to a place where there's been a big step forward while largely ignoring how it relates to human-level thought. (After all, we got powered flight by largely ignoring how birds fly.) In fact, cognitive science largely split off as a separate field but there's a school of thought that algorithms and big data are only going to take things to a certain point and you're not going to get to things like fully autonomous vehicles under general conditions without different approaches that involve better understanding the physical world.
  - bluetwo 9y ago
    Actually, previous to the Wright brothers, the dominant theories of flight involved studying how birds flew and theorizing what parts were important and trying to replicate them. You could get a degree in this area of study. I do agree that different approaches are needed. In the case of flight it was the realization that the airfoil is important and the rest is window dressing.
    - jv22222 9y ago
      Am I right in thinking the study of fish/sharks and fluid dynamics helped with the breakthrough?
      - mgraczyk 9y ago
        No, it was bicycles that guided the intuition that control should be based on roll rather than rudder yaw. That was the key innovation that the Wrights contributed at the time of their first flight.
bobsil1 9y ago
Hand-engineering and hand-assembling evolved components will fail, way too complex. We struggle with making bug-free word processors. This will only be solved by fully evolving AGI, or some other form of automated search of high-dimensional design space.
SomeStupidPoint 9y ago
I don't think you're wrong that AGI will have (something like) those components in its inception. I just disagree that we have to intentionally create them -- I think there are a lot of places we can experimentally bruteforce the implementation instead of having a solid theoretical understanding if our only goal is creating an AGI. I don't think we need any theoretical breakthroughs, per se, but rather, a lot of computing power and time. No one seems willing to take a million nodes and run them for a decade to brute force some of the mechanisms -- everyone wants results on the short timelines that grants or Wall St operate on. I get why, but that bias towards quantifiable short-term gains fundamentally limits the search algorithms we're implementing in our quest for AI, and likely means that we won't get there in a meaningful way, because intentionally implementing the necessary requirements is a fool's errand of complexity. In short: we're not smart enough to do the simple, but slow thing, so we're trying the highly complex one with demonstrable incremental results. I expect this approach to continuously fail to develop AGI even while it demonstrates results on discrete problems.
- aeleos 9y ago
  I think you are right that we could, in theory experimentally brute force intelligence. But if we try and do that, we are basically trying to simulate an entire universe and wait for intelligence. But the amount of computing power that that would require is beyond even what we can imagine. And even if we could find some ways of simplifying it, it still seems like a waste of our intelligence. Like when it got cold for humans, rather than just dying and waiting until they evolved fur, they just made clothes. Evolution is inherently able to do great things but is very inefficient. We have the power to reverse engineer nature and use that knowledge to our advantage, and if our goal is to create general AI, then not trying to understand the one sitting in everyone heads and just trying to throw more computing behind it just seems like a waste of a tool that has allowed us to get this far.
  - SomeStupidPoint 9y ago
    I'll preface this by saying that my views don't necessarily represent a mainstream position here: Humans already accidentally invented AI at least a few times, sort of. To talk meaningfully about this topic, I think it's better to step away from the topic of "intelligence" and instead speak about "reasoning ability". It's still fuzzy, but seems to sidestep a lot of the "can submarines swim" arguments. Things like societies, governments, corporations, economies, etc all show similar features to machine learning algorithms and reasoning animals in their ability to dynamically problem solve -- the main difference is that they run on human computers instead of electronic computers. (Or photonic, quantum, etc computers.) Well, in modern times, they're actually sort of cyborgs, but that's a side issue. The quest for an AGI is more about translating the notion of a corporation (or government, etc) meaningfully into fully electronic computer terms than it is in replicating the human mind. We know enough about the supra-structure of the corporation to meaningfully bound the search, and it's much more prudent to spend a decade on bruteforcing each component of a corporation than it is to navel gaze about reinventing minds. The truth is that human brains (or animal brains) aren't the only general intelligence that we know -- so reinventing them isn't the only path to AGI, though it's a worthwhile project in its own right. I just worry that many of the meaningful contributions of the people working on reinventing minds will be lost on the birth of the actual first AGI through corporate cyberization. (And that many of us will fail to recognize corporate cyberization for what it is -- an AGI. Any (successful) corporation which can act fully by algorithm must be a general intelligence.)
mythrwy 9y ago
Background: Some guy on the Internet with an unrelated Bachelors degree he didn't study that hard for and has picked up a little Python and Javascript somewhere along the way. I know of no biological forms that have evolved wheels. But wheels have turned out to be a hell of a lot faster than fins or legs. I see no reason "intelligence" has to follow the pathways or limitations of biology or neurology at all. Although certainly it may be a place to look for some ideas.
- npgatech 9y ago
  Fair point, but you are omitting the fact that brain itself (that invented the wheel) was evolved through biological means. That said, evolutionary optimization is generally considered unreliable(convergence) and inefficient.
  - mythrwy 9y ago
    The wheel existed before the mind did. It was just sitting there waiting to be noticed :)
    - pookeh 9y ago
      Not for the function the brain made a use of it.
- skummetmaelk 9y ago
  Wheels are useless if you don't have roads. EDIT: Please explain your downvotes here? Don't you people see that this, coupled with there being no useful intermediate stage before developing a full wheel, prevents evolution from "inventing" it.
  - jamiek88 9y ago
    Well that's not true at all. See: Wheelbarrow; amongst many wheeled devices I can think of immediately.
aeleos 9y ago
What do you think the best way of attacking these problems is? It seems to me that we aren't even at the point where technology can do much to help us (in terms of recreating what we know about the brain) and we just need to understand the brain better.
- ghaff 9y ago
  "Just" that. Maybe. The problem is that, at the risk of trivializing a lot of good research, science related to the brain/how we think/etc. has seen a whole lot more money spent and the effort of smart people expended than it has seen useful results. There are still debates going on about a lot of aspects of learning and other aspects of cognitive science that were probably already happening when computers were made from vacuum tubes or even earlier.
moultano 9y ago
>Every time a major advance is made in AI the computational neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!" Can you name one time that has happened? I can't think of any.
- dragandj 9y ago
  Hindsight is 20/20 there.
- eli_gottlieb 9y ago
  Bayesian brain theories of perception. The neuroscientists and cognitive scientists are still waiting for the AI theorists to stop obsessing over that one paper with the cat's visual cortex and get off the deep learning train of sacrificing correctness for cheapness to compute.
  - return0 9y ago
    this is not a case that happened
    - eli_gottlieb 9y ago
      As a matter of fact, the mainstream AI community hasn't adopted probabilistic programming, but it does exist.
      - aoeuasdf1 9y ago
        Feel free to link to any Kaggle blog posts showing victories using probabilistic programming. Is your opinion independent of evidence?
      - return0 9y ago
        you re going way off topic . this is about AI discoveries which in retrospect were found to have been discovered by neuroscience. you 're complaining that a potential theory (which is only partially supported by evidence) might one day hypothetically be found useful in AI.
trott 9y ago
> Every time a major advance is made in AI the computational neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!". I'm not sure which advances you are talking about, but the modern successes of Deep Learning are primarily due to backpropagation and convolution. Neither of the two is considered to be biologically plausible. Some people are actively trying to come up with alternatives to backpropagation that would be biologically plausible though. See for example: Bengio et al., "Towards Biologically Plausible Deep Learning" https://arxiv.org/abs/1502.04156 https://arxiv.org/abs/1502.04156
- p1esk 9y ago
  Actually convnets were inspired by Fukushima's Neocognitron, which was itself inspired by visual cortex.
  - trott 9y ago
    > Actually convnets were inspired by Fukushima's Neocognitron, which was itself inspired by visual cortex. That doesn't contradict what I wrote. ConvNets require the synchronization of weights between neurons, which is not considered to be biologically plausible. Some aspects of the architecture (the receptive fields, in this case) may well be plausible, with the complete architecture still implausible.
    - rdlecler1 9y ago
      You're focusing too much on the implementation details and assuming that because they're different that it's not equivalent. The secret sauce here is that network topology (and threshold rules), not the implementation details, are largely what determine the functional properties of that network. Show an electrical engineer the circuit diagram of a 4 bit adder and they'll know it's function immediately. Artificial Neural Networks and Artificial Gene Regulatory Networks are the same. The problem with ANNs is that we fail to see the circuitry driving the function because we assume that every w_ij != 0 is functionally relevant. Once you start to strip away non-spurious interactions you start to see topological patterns. I did a lot of work on this problem -- this is not an uneducated guess.
laxatives 9y ago
Don't know why you feel so compelled to give your background while writing an anonymous comment on the internet. It only makes me suspect your argument doesn't hold up without some expectation to respect your authority. All the theory in the world is irrelevant until it you can apply it to predict something new -- its almost trivial to create a theory that describes some phenomenon accurately, after-the-fact. See the tremendous bifurcation in physics with regards to string theory. Similarly there are loads of academics in various fields claiming to have been the origins of artificial intelligence.

heimatau 9y ago

Hmm, reading your experience clarifies to me (a Maths guy) why current deep learning efforts are producing real fruit from their labor. Linear Regression and Vector/dot products are just constructs to re-create the evolutionary biology seen in nature. When looking at the slow process of evolution, it's the nuances that matter and over time, they compound exponentially. Pretty neat. By breaking down a task into abstract constructs and sifting problems out with a sigmoid function.

eggie 9y ago

I'd love to read your thesis. Would you please a link?

Jedi72 9y ago

Can you link to some of your work, this is very interesting stuff

return0 9y ago

> neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!". Interested to hear about these. I can honestly not remember any.

ghaff 9y ago
There's been this tension over whether achieving apparent human levels of artificial intelligence requires thinking like and following processes associated with human learning and intelligence pretty much forever. Today, computational capability and data have gotten us to a place where there's been a big step forward while largely ignoring how it relates to human-level thought. (After all, we got powered flight by largely ignoring how birds fly.) In fact, cognitive science largely split off as a separate field but there's a school of thought that algorithms and big data are only going to take things to a certain point and you're not going to get to things like fully autonomous vehicles under general conditions without different approaches that involve better understanding the physical world.
- bluetwo 9y ago
  Actually, previous to the Wright brothers, the dominant theories of flight involved studying how birds flew and theorizing what parts were important and trying to replicate them. You could get a degree in this area of study. I do agree that different approaches are needed. In the case of flight it was the realization that the airfoil is important and the rest is window dressing.
  - jv22222 9y ago
    Am I right in thinking the study of fish/sharks and fluid dynamics helped with the breakthrough?
    - mgraczyk 9y ago
      No, it was bicycles that guided the intuition that control should be based on roll rather than rudder yaw. That was the key innovation that the Wrights contributed at the time of their first flight.

bobsil1 9y ago

Hand-engineering and hand-assembling evolved components will fail, way too complex. We struggle with making bug-free word processors. This will only be solved by fully evolving AGI, or some other form of automated search of high-dimensional design space.

SomeStupidPoint 9y ago

I don't think you're wrong that AGI will have (something like) those components in its inception. I just disagree that we have to intentionally create them -- I think there are a lot of places we can experimentally bruteforce the implementation instead of having a solid theoretical understanding if our only goal is creating an AGI. I don't think we need any theoretical breakthroughs, per se, but rather, a lot of computing power and time. No one seems willing to take a million nodes and run them for a decade to brute force some of the mechanisms -- everyone wants results on the short timelines that grants or Wall St operate on. I get why, but that bias towards quantifiable short-term gains fundamentally limits the search algorithms we're implementing in our quest for AI, and likely means that we won't get there in a meaningful way, because intentionally implementing the necessary requirements is a fool's errand of complexity. In short: we're not smart enough to do the simple, but slow thing, so we're trying the highly complex one with demonstrable incremental results. I expect this approach to continuously fail to develop AGI even while it demonstrates results on discrete problems.

aeleos 9y ago
I think you are right that we could, in theory experimentally brute force intelligence. But if we try and do that, we are basically trying to simulate an entire universe and wait for intelligence. But the amount of computing power that that would require is beyond even what we can imagine. And even if we could find some ways of simplifying it, it still seems like a waste of our intelligence. Like when it got cold for humans, rather than just dying and waiting until they evolved fur, they just made clothes. Evolution is inherently able to do great things but is very inefficient. We have the power to reverse engineer nature and use that knowledge to our advantage, and if our goal is to create general AI, then not trying to understand the one sitting in everyone heads and just trying to throw more computing behind it just seems like a waste of a tool that has allowed us to get this far.
- SomeStupidPoint 9y ago
  I'll preface this by saying that my views don't necessarily represent a mainstream position here: Humans already accidentally invented AI at least a few times, sort of. To talk meaningfully about this topic, I think it's better to step away from the topic of "intelligence" and instead speak about "reasoning ability". It's still fuzzy, but seems to sidestep a lot of the "can submarines swim" arguments. Things like societies, governments, corporations, economies, etc all show similar features to machine learning algorithms and reasoning animals in their ability to dynamically problem solve -- the main difference is that they run on human computers instead of electronic computers. (Or photonic, quantum, etc computers.) Well, in modern times, they're actually sort of cyborgs, but that's a side issue. The quest for an AGI is more about translating the notion of a corporation (or government, etc) meaningfully into fully electronic computer terms than it is in replicating the human mind. We know enough about the supra-structure of the corporation to meaningfully bound the search, and it's much more prudent to spend a decade on bruteforcing each component of a corporation than it is to navel gaze about reinventing minds. The truth is that human brains (or animal brains) aren't the only general intelligence that we know -- so reinventing them isn't the only path to AGI, though it's a worthwhile project in its own right. I just worry that many of the meaningful contributions of the people working on reinventing minds will be lost on the birth of the actual first AGI through corporate cyberization. (And that many of us will fail to recognize corporate cyberization for what it is -- an AGI. Any (successful) corporation which can act fully by algorithm must be a general intelligence.)

mythrwy 9y ago

Background: Some guy on the Internet with an unrelated Bachelors degree he didn't study that hard for and has picked up a little Python and Javascript somewhere along the way. I know of no biological forms that have evolved wheels. But wheels have turned out to be a hell of a lot faster than fins or legs. I see no reason "intelligence" has to follow the pathways or limitations of biology or neurology at all. Although certainly it may be a place to look for some ideas.

npgatech 9y ago
Fair point, but you are omitting the fact that brain itself (that invented the wheel) was evolved through biological means. That said, evolutionary optimization is generally considered unreliable(convergence) and inefficient.
- mythrwy 9y ago
  The wheel existed before the mind did. It was just sitting there waiting to be noticed :)
  - pookeh 9y ago
    Not for the function the brain made a use of it.
skummetmaelk 9y ago
Wheels are useless if you don't have roads. EDIT: Please explain your downvotes here? Don't you people see that this, coupled with there being no useful intermediate stage before developing a full wheel, prevents evolution from "inventing" it.
- jamiek88 9y ago
  Well that's not true at all. See: Wheelbarrow; amongst many wheeled devices I can think of immediately.

aeleos 9y ago

What do you think the best way of attacking these problems is? It seems to me that we aren't even at the point where technology can do much to help us (in terms of recreating what we know about the brain) and we just need to understand the brain better.

ghaff 9y ago
"Just" that. Maybe. The problem is that, at the risk of trivializing a lot of good research, science related to the brain/how we think/etc. has seen a whole lot more money spent and the effort of smart people expended than it has seen useful results. There are still debates going on about a lot of aspects of learning and other aspects of cognitive science that were probably already happening when computers were made from vacuum tubes or even earlier.

moultano 9y ago

>Every time a major advance is made in AI the computational neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!" Can you name one time that has happened? I can't think of any.

dragandj 9y ago
Hindsight is 20/20 there.
eli_gottlieb 9y ago
Bayesian brain theories of perception. The neuroscientists and cognitive scientists are still waiting for the AI theorists to stop obsessing over that one paper with the cat's visual cortex and get off the deep learning train of sacrificing correctness for cheapness to compute.
- return0 9y ago
  this is not a case that happened
  - eli_gottlieb 9y ago
    As a matter of fact, the mainstream AI community hasn't adopted probabilistic programming, but it does exist.
    - aoeuasdf1 9y ago
      Feel free to link to any Kaggle blog posts showing victories using probabilistic programming. Is your opinion independent of evidence?
    - return0 9y ago
      you re going way off topic . this is about AI discoveries which in retrospect were found to have been discovered by neuroscience. you 're complaining that a potential theory (which is only partially supported by evidence) might one day hypothetically be found useful in AI.

trott 9y ago

> Every time a major advance is made in AI the computational neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!". I'm not sure which advances you are talking about, but the modern successes of Deep Learning are primarily due to backpropagation and convolution. Neither of the two is considered to be biologically plausible. Some people are actively trying to come up with alternatives to backpropagation that would be biologically plausible though. See for example: Bengio et al., "Towards Biologically Plausible Deep Learning" https://arxiv.org/abs/1502.04156 https://arxiv.org/abs/1502.04156

p1esk 9y ago
Actually convnets were inspired by Fukushima's Neocognitron, which was itself inspired by visual cortex.
- trott 9y ago
  > Actually convnets were inspired by Fukushima's Neocognitron, which was itself inspired by visual cortex. That doesn't contradict what I wrote. ConvNets require the synchronization of weights between neurons, which is not considered to be biologically plausible. Some aspects of the architecture (the receptive fields, in this case) may well be plausible, with the complete architecture still implausible.
  - rdlecler1 9y ago
    You're focusing too much on the implementation details and assuming that because they're different that it's not equivalent. The secret sauce here is that network topology (and threshold rules), not the implementation details, are largely what determine the functional properties of that network. Show an electrical engineer the circuit diagram of a 4 bit adder and they'll know it's function immediately. Artificial Neural Networks and Artificial Gene Regulatory Networks are the same. The problem with ANNs is that we fail to see the circuitry driving the function because we assume that every w_ij != 0 is functionally relevant. Once you start to strip away non-spurious interactions you start to see topological patterns. I did a lot of work on this problem -- this is not an uneducated guess.

laxatives 9y ago

Don't know why you feel so compelled to give your background while writing an anonymous comment on the internet. It only makes me suspect your argument doesn't hold up without some expectation to respect your authority. All the theory in the world is irrelevant until it you can apply it to predict something new -- its almost trivial to create a theory that describes some phenomenon accurately, after-the-fact. See the tremendous bifurcation in physics with regards to string theory. Similarly there are loads of academics in various fields claiming to have been the origins of artificial intelligence.