From: Hector Zenil (firstname.lastname@example.org)
Date: Fri Nov 23 2007 - 23:53:56 MST
I did one of my theses on Recurrent Analog Neural Networks (ARNN) and
gave a talk at the University of Liverpool two years ago, the
proceedings were recently published by World Scientific with my
paper under the title "On the possible Computational Power of the
Human Mind". An arXiv version is available online:
and the presentation from the conference:
ARNNs are basically neural networks, first analyzed by Siegelmann and
Sontag, able to allow arbitrary real numbers as weights.
Traditionally, neural networks have been considered to have rational
or computable weights. From Kleene we know that neural networks with
integer weights are equivalent in computational power to finite
automata. Siegelmann made several contributions in her seminal book
such as the construction of a very small universal neural network with
only 9 neurons (published before in joint with Maurice
Margenstern). She also gave several interesting computational
complexity results related to the ARNN's in general and provided a
proof for neural networks with rational numbers as equivalent to
Turing machines in computational power. Then extended the model to
allow any arbitrary real number that would then automatically allow
the ARNN to compute beyond the Turing limit (e.g. put as weight the
value of $\Omega$ the halting set coded as a real number). Of course
one question is from where one can get that non-computable number in
order to put it in the net configuration to reach hypercomputation
(all known artificial neural networks run basically over Turing
machines), although if the system already has it (e.g. one already
operating in nature or in our head) then it automatically would reach
hypercomputation. But from it one cannot draw any conclusion about its
feasibility. Martin Davis's excellent article , "The Myth of
Hypercomputation," clearly articulates several criticisms, including
jumps to conclusions from Siegelmann's work.
Hava Siegelmann is not claiming that the human brain is an ARNN, we
suggest that it can be studied through an ARNN model though, because
of its generality. Of course a brain with a neural network with
arbitrary real number would go beyond the Turing limit, but that is
unlikely, even when there are other elements traditionally taken as
"continuous" variables, such as neural spikes, also covered in .
 Hector Zenil & Francisco Hernandez-Quiroz (2007), Worldviews, On
the possible Computational Power of the Human Mind. Worldviews,
Science and Us: Philosophy and Complexity by Carlos Gershenson, et. al
(ed.), World Scientific, pp. 315-337.
from Amazon: http://www.amazon.com/Worldviews-Science-Us-Philosophy-Complexity/dp/9812705481/ref=sr_1_1?ie=UTF8&s=books&qid=1195886116&sr=8-1Yf
 Hava T. Siegelmann (1998), Neural Networks and Analog Computation:
Beyond the Turing Limit, Springer.
 Hava T. Siegelmann, Maurice Margenstern (1999), Nine switch-affine
neurons suffice for Turing universality. Neural Networks 12(4-5):
 Davis, M. (2004) The Myth of Hypercomputation. Alan Turing: Life
and Legacy of. a Great Thinker, Teuscher, C. (ed.) Springer, pp.
On Nov 24, 2007 6:39 AM, <email@example.com> wrote:
> Is a human brain or biological neural network superior to a computer
> or Turing machine?
> Has this question ever been answered? There are some scientists that
> claim that the brain has more power than any Turing machine.
> One example would be the following book:
> Hava T. Siegelmann
> Neural Networks and Analog Computation: Beyond the Turing Limit
-- Hector Zenil-Chavez firstname.lastname@example.org Université de Lille I (Laboratoire d'Informatique Fondamentale) Université Pantheon-Sorbonne -Paris 1- (IHPST) -------------------------------- zenil.mathrix.org animaexmachina.com --------------------------------- Fondation Suisse Cité Internationale Universitaire de Paris 7, bd Jourdan - chr. 114 75014 Paris France --------------------------------------------------------------
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:01 MDT