nettalk.readme

著名的神经网络工具箱

=============================================================================
	README file for the example files nettalk.xxx
=============================================================================


Description:	The NETtalk network, a famous feedforward-network
============	

This network is a lookalike to the nettalk network, a famous
feedforward-network, developed by T.J. Sejnowski and C.R. Rosenfeld:

T.J. Sejnowski and C.R. Rosenfeld:NETtalk: a parallel network that
learns to read aloud, The Johns Hopkins Univ. EE and
Comp. Sc. Technical Report JHU/EECS-86/01, 32 pp., also in:
J.A. Anderson, E. Rosenfeld, Neurocomputing, Foundations of Resarch,
MIT Press, 1988

The nettalk network given here is NOT the fully trained NETtalk
network, but a lookalike with the same topology (same number of units,
same connectivity, same number of weights). We use this network mainly
for benchmarking, since the nettalk network has become a de facto
standard benchmarking network.


Pattern-Files:	nettalk.pat
==============

The pattern file nettalk.pat is NOT the original NETtalk pattern file
but also a smaller lookalike with a similar statistical distribution
of input bits set to '1'.  See the paper cited above for the real
NETtalk data and input/output coding.

This pattern file, which we use both for training and test of 'raw
speed' backpropagation benchmarks of SNNS on different machines,
operating systems and compilers, consists of only 200 input patterns,
whereas the original NETtalk data consists of 1000 input patterns
taken from an even larger dictionary of approx. 20.000 input patterns.
Since convergence of the algorithm doesn't matter in backpropagation
implementation speed benchmarks this is no problem.


Network-Files:	nettalk.net
==============

The nettalk network consists of
	203 input units
	120 hidden units
	 26 output units
All layers are fully connected without shortcut connections.


Config-Files:	nettalk.cfg
=============

Since this is no 'real' application and data, it is not too
interesting to view the network. The configuration file nettalk.cfg is
for those who want to view it anyway.


Hints:
======

We usually call nettalk.net from the tool 'netperf', a benchmarking
version of SNNS without graphical user interface but with the full and
current SNNS kernel. It is supplied with SNNS in the tools directory.
We usually train and test nettalk for 10 cycles. Since convergence is
not important, the learning parameters can be set to arbitrary (small)
values or even be set to 0.

Performance numbers for SNNS supplied by external testers should best
be obtained the same way. If you send us the training and recall speed
of SNNS on your machine for this network, please specify in detail the
SNNS version, machine model (maker, type), main memory size, operating
system version, compiler and compiler optimization level at
installation of SNNS.


=============================================================================
	End of README file
=============================================================================