On AI and the brain

AGI [artificial general intelligence] must be possible. That is because of a deep property of the laws of physics, namely the universality of computation. It entails that everything that the laws of physics require physical objects to do can, in principle, be emulated in arbitrarily fine detail by some program on a general-purpose computer, provided it is given enough time and memory.
--David Deutch on his blog at the Guardian, 3 October 2012.

Why is this nonsense so loudly proclaimed by those who prefer to conceive of the brain, indeed life on Earth,  indeed the universe, as a mere computer?  Where is the fire in the equations? A computer simulation of a nuclear explosion, be it oh so accurate, is not a nuclear explosion! And those who claim that simulating intelligence via a computer would be intelligence unlike an explosion simulation being an explosion, since the output would be intelligent in content, are begging the question, since it would only be truly intelligent as a simulation if the simulation succeeded in being intelligent, which is the question at issue!

Let's quote Selim Akl about the fallacy of assuming that the finitely computable algorithms that are currently within reach of microprocessor technology will work for intelligence:

Non-universality in computation: Given n spatially and temporally connected physical variables, X1, X2, ..., Xn, where n is a positive integer, and a function F(X1, X2, ..., Xn) of these variables, no computer can evaluate F for any arbitrary n, unless it is capable of an infinite number of operations per time unit.
Note that F is readily computable by a machine M capable of exactly n operations per time unit. However, this machine cannot compute F when the number of variables is n+1. While a second machine M' capable of n+1 operations per time unit can now compute the function F of n+1 variables, M' is in turn defeated by a function of n+2 variables. This continues forever.
This point deserves emphasis. While the function F(X1, X2, ..., Xn+1) = F1(X1), F2(X2), ..., Fn+1(Xn+1) is easily computed by M', it cannot be computed by M. Even if given infinite amounts of time and space, machine M is incapable of simulating the actions of M'. Furthermore, machine M' is in turn thwarted by F(X1, X2, ..., Xn+2), a function computable by a third machine M''. This continues indefinitely. Therefore no computer is universal if it is capable of exactly T(i) operations during time unit i, where i is a positive integer, and T(i) is finite and fixed once and for all (for it will be faced with a computation requiring V(i) operations during time unit i, where V(i) > T(i) for all i).
Examples of such function F occur in:
1. Computations with time-varying variables: The variables, over which the function is to be computed, are themselves changing with time.
2. Computations with time-varying computational complexity: The computational complexity of the function to be computed is itself changing with time.
3. Computations with rank-varying computational complexity: Given several functions to be computed, and a schedule for computing them, the computational complexity of a function depends on its position in the schedule.
4. Computations with interacting variables: The variables of the function to be computed are parameters of a physical system that interact unpredictably when the system is disturbed.
5. Computations with global mathematical constraints: The function to be computed is over a system whose variables must collectively obey a mathematical condition at all times.
6. Computations with uncertain time constraints: There is uncertainty with regards to the input (when and for how long are the input data available), the calculation (what to do and when to do it), and the output (the deadlines are undefined at the outset); furthermore, the function that resolves each of these uncertainties itself has demanding time requirements.

--Akl, SG, in http://research.cs.queensu.ca/Parallel/projects.html;  see also Akl, S.G., The Myth of Universal Computation, in Parallel Numerics ’05,M. Vajterňásic, R. Trobec, P. Zinterhof, A. Uhl (Eds.), chapter 7.

Especially, consider #4 above, since that is how the brain works. This is not to say that a breakthrough in intricacy of nanotechnology could not make AI a reality.  But Moore's Law alone acting on our current microprocessor based computer tech cannot achieve human-equivalent general purpose intelligence.

And right now, I'd be happy just with computer dictation of my notes that made less than 1 error per paragraph, Mr. Kurzweil. The computer-generated typos do creep in, and our current braindead (back to AI here) EMR is security-restricted to disallow corrections of finished notes.


No comments:

Post a Comment

Happy Thanksgiving!

Let's even be grateful that we can be grateful.