Tech Report Available

[Hideyuki Kato]

The following techinical report  is now available in neuroprose:


            Power law in the performance of the human memory 
                                and
              a simulation with a neural network model               


Tatsuhiro Nakajima,  Nobuko Fuchikami
Department of Physics, Tokyo Metropolitan University
1-1 Minami-Osawa, Hachioji, Tokyo 192-03, Japan

Hideyuki Cateau
Department of Physics, University of Tokyo
Bunkyo-ku, Tokyo 113, Japan

Hiroshi Nunokawa
National Laboratory for High Energy Physics (KEK)
Tsukuba-shi, Ibaraki 305, Japan


This paper will appear in proceedings of ISKIT'92.
The report number of this paper is TMUP-HEL-9203, TU-611, KEK-TH-334
or KEK preprint 92-40.


Abstract

We show that a learning pace of the back propagation model is 
described by a power law 
with high precision.  Interestingly the 
same power law was found out in the human memory by a psychologist 
in the past.   
Therefore our result provides a quantitative evidence that the back 
propagation  model, though it is simple, surely shares 
 some essential structure with the human brain.  In proceeding the 
discussion we make out a novel memory model.  
This model naturally avoids the notable difficulty of 
the back propagation network that the learning  of it  is very 
sensitive to the initial condition.  

   * Hard copy request is not available, sorry*

Instructions for obtaining by anonymous ftp:

%ftp archive.cis.ohio-state.edu
Name: anonymous
Password:neuron
ftp>bin
ftp>cd pub/neuroprose
ftp>get nakajima.power.tar.Z
ftp>quit
%uncompress nakajima.power.tar.Z
%tar xvfo nakajima.power.tar


Hideyuki Cateau

Department of Physics, 
University of Tokyo,
Hongo, Bunkyo-ku, Tokyo, 113 Japan