Re: [ai-philosophy] Provable generality?
From: Bill Modlin <wdmodlin@...>
Sent: Mon, June 28, 2010 9:55:44 PM
Subject: Re: [ai-philosophy] Provable generality?
In your casual description of training the system for the early stages of
processing, it sounds as though you assume that these algorithms can be
developed once and frozen, that while it may take "a lot of training" to set it
up in the first place, once it is done it is done.
In the human brain, it appears the early stages retain some plasticity even into
adulthood, that feedback from higher levels of cognitive processing can modify
the details. We can learn new feature detection algorithms and tune our
discrimination of existing features. In other words, the training is never
really finished for any part of the system. I'm concerned that even "adaptive"
Levin search is not suitable for the kind of continuous iterative tuning of
algorithms that is supported by a network of adjustable connections and
weights. The granularity of adaptive changes is too coarse, the computational
work required for an adaptation is much more than is required for a local
I have similar granularity-related concerns about the HTM methods also under
I want to point-out that a "changing program" is a ill definition a program
never change by definition, a changing program is not a program ... . If you are
watching a changing program you are not watching the real program.
And again a program can be "not written somewhere" you can have a system without
an explicit program .
So a brain change its network etc... this means that you can not reproduce
exactly the brain to reproduce "the changing brain" functionality.
Using an Inverse Levin Search what you build is the correct ultimate program ,
also you train a human ( in a hypothetical exeriment) .
On Fri, Jul 2, 2010 at 10:58 AM, Denis <dnsflex@...> wrote:>
>From: Bill Modlin <wdmodlin@...>>Subject: Re: [ai-philosophy] Provable generality?
>>--- On Thu, 7/1/10, Denis <dnsflex@...> wrote:No , I am not missing that points .
>> Using an Inverse Levin Search what you build is the correct ultimate program
>I think you are missing important points.
The problem of an Inverse Levin Search is not "changing of new information", the
problem is an absurd requirement of resources .
I try to explain :
If you have an initial training set T1 you can run the I.L.S. so you find a
program P1 now if you have a new training set T2 you can run again the I.L.S for
the training T1+T2 you have 2 possibilities for the result 1) the solution is
again P1 2) the solution is a program P2 .
The second case mean that was impossible to find P2 without more information! To
find P2 you need more information than T1 ! Find P1 is the best you can do with
only T1 .I don't understand what "inverse" levin search is here.
OK, this is only theoretical becouse there are not enough resource to do
something like this.But this is what the inverse search do!
>For one thing, situations change and new information arrives all the time. So
>even if your inductive engine could build the "correct ultimate program" based
>on all the information available today, tomorrow you might need to modify it.
>For another, it is computationally infeasible to take into account all the
>information available all at once. You cannot really build the "correct
>You must instead spend computational energy developing
>simplified views of the data, abstracting what seem likely to be important
>aspects, features and relationships for further examination.
that is what levin search does in fact, but inverse levin search?> Once you begin toagain this is what the inverse search do.
>see how it fits together, you may find reasons to modify the abstraction
>algorithms, to retain more or less detail about various features or to look for
Yes we are always in approximations , the problem is in general uncomputable.
> At any point in time you can have at best an approximation to the
To be exact, it is semi-computable. It is computable in the limit.Best,
Eray Ozkural, PhD candidate. Comp. Sci. Dept., Bilkent University, Ankara