paper: dynamics of learning in twolayered networks

[Michael Biehl]

FTP-host:       ftp.physik.uni-wuerzburg.de 
FTP-filename:   /pub/preprint/WUE-ITP-96-003.ps.gz

The following paper is now available via anonymous ftp:
(See below for the retrieval procedure)

------------------------------------------------------------------
"Transient dynamics of on-line learning in two-layered neural networks"

Michael Biehl, Peter Riegler, and Christian W"ohler

Ref. WUE-ITP-96-003     

			  Abstract
The dynamics of on-line learning in neural networks with continous units
is dominated by plateaus in the time dependence of the generalization error.
Using tools from statistical mechanics, we show for a soft committee machine
the existence of several fixed points of the dynamics of learning that give
rise to complicated behavior, such as cascade--like runs through different
plateaus with a decreasing value of the corresponding generalization error.
We find learning-rate dependent phenomena, such as splitting and disappearing
of fixed points of the equations of motion. The dependence of plateau lengths
on the initial conditions is described analytically, and simulations confirm
the results.   

---------------------------------------------------------------------

Retrieval procedure:

     unix> ftp  ftp.physik.uni-wuerzburg.de
     Name: anonymous  Password: {your e-mail address}
     ftp>  cd pub/preprint
     ftp>  get WUE-ITP-96-003.ps.gz           (*)   
     ftp>  quit
     unix> gunzip WUE-ITP-96-003.ps.gz
e.g. unix> lp WUE-ITP-96-003.ps                     [15 pages]             

(*) can be replaced by "get WUE-ITP-96-003.ps". The file will then
    be uncompressed before transmission (slow!). 
_____________________________________________________________________ 


Michael Biehl     
Institut fuer Theoretische Physik              
Julius-Maximilians-Universitaet Wuerzburg
Am Hubland
D-97074 Wuerzburg 

email:  biehl at physik.uni-wuerzburg.de
Tel.:   (+49) (0)931 888 5865   
          "      "    "  5131
Fax :   (+49) (0)931 888 5141