In my last post on artificial intelligence, I deferred to Robin Hanson on his claim that strong AI achieved through brain emulation is far easier than weak AI based on manual programming in order to focus on my core claim that weak AI is preferable for behavioral reasons. Robin worked in the 80s as an AI researcher at Lockheed and in the early 90s as a researcher for NASA, so he clearly has more expertise than I do on the question of what is feasible in the field.
Though I doubt Robin’s claim less than I would if he weren’t making it, I doubt it nonetheless. In part this is because of pessimism that we will ever be able to fully emulate the biology of the brain. As Timothy Lee argues, biological processes are bottom-up; computers are designed top-down.
But more importantly, after today I’m actually pretty impressed with the state of weak AI. Via Adam Gurri, I found Engadget’s post on an IBM supercomputer that seems to perform quite well against human competitors at Jeopardy. Watch the videos embedded in the article; they are surprising and delightful.
After watching that, how can anyone think that weak AI is futile?
Do you know where neural networks stand vis à vis weak or strong AI? Neural networks are predictive algorithms that are “trained” using data sets and are conceptually very different from regressions. I just learned about the concept a few weeks ago, but it seems to be a step in the direction of “brain emulation.”
The distinction between strong and weak AI is not in the method of computation but in the generality of the intelligence. Something is a strong AI if it is capable of performing the same range of mental tasks that humans can perform (some people also associate strong AI with subjective experience and sentience). In contrast, weak AI is more limited in that it is tuned to solve a particular sort of problem. Robin’s claim as I understand it is that we’re better off leapfrogging attempts at weak AI (including those through neural networks) and instead using whole brain emulation to achieve strong AI.
Pingback: Overcoming Bias : Signal Processors Decouple
“biological processes are bottom-up; computers are designed top-down.”
And ?
We live in the same universe, both are physical systems…
I suggest you read the Church-Turing thesis, you can skip the math if you are so inclined but make sure to read about the conclusions and implications…
Then read David Duetsch, “bottom up” and “top down” is just silly made up rationalizations.
A computer is not some inferior “top-down” system. It is the most remarkable system of them all, it can predict and evolve the state of any physical system possible, even magic “bottom-up” ones. :)
Proper Dave,
I’m not saying it’s literally impossible, I understand computability, etc.
@Proper Dave, I don’t think you know what you’re talking about. The Church-Turing thesis tells us what’s computable given infinite time and resources. Even granting that the brain can be modeled with a Turing Machine, the C-T Thesis doesn’t tell us whether such a model is mathematically tractable. Computers can be simulated efficiently because human beings designed them in a way that made that inevitable. Natural systems weren’t designed by anyone, so there’s no particular reason to think we can simulate them efficiently.
I find your use of “bottom up” and “top down” puzzling then, they are fundamentally “incompatible” no? In different “realms” or whatever.
I don’t know if you’re replying to me or to Tim, but speaking only for myself, the point is that emergent, bottom-up processes frequently generate such staggering complexity that in practice they cannot be understood well enough to virtualize. Consciously-designed, top-down phenomena are limited in their complexity by necessity, because a designer or designers must understand them well enough to implement them in the first place.
The point is not that bottom-up processes are fundamentally exempt from being computable in a theoretical sense, but that there is a binding resource constraint in actually computing some of them. We do not have “infinite tape.”
Yes that’s true, we just may be to dumb to understand that. I find it very difficult to believe though you can just run an ab initio simulation and tweak, it may take hundreds or even thousands of years, but you just may find the essential “higher order” functions and discard all the unnecessary lower order stuff. What will gum up the works is if there is some fundamental roadblock in reaching theoretical computational efficiency and we hit it soon.
“Top down” processes is very new only about 200 years for the scientific method and engineering principles, natural selection had about 3 billion years, they are both evolutionary processes, so I think the jury is still out about the inferiority of the “top-down” methods.
You don’t need infinite tape, no physical system is going to be intractable or produce a combinatorial explosion, ironically intractable problems are abstract and complex stuff dreamed up by the inferior “top-down” methods.