TR available

[Xie Yun]

The following technical report is placed in the neuroprose archive:

\title{Training Algorithms for Limited Precision Feedforward Neural Networks }
\author{Yun Xie \thanks{Permanent address:
        Department of Electronic Engineering,
        Tsinghua University,
        Beijing 100084, P.R.China}
        \hskip 30pt Marwan A. Jabri\\
        \\
        Department of Electrical Engineering\\
        The University of Sydney\\
        N.S.W. 2006, Australia}
\date{}
\maketitle

\begin{abstract}
A statistical quantization model is used to analyze the effects of
quantization on the performance and the training dynamics of a feedforward
multi-layer neural network implemented in digital hardware.  The analysis
shows that special techniques have to be employed to train such networks
in which each variable is represented by limited number of bits in  fixed
point format.  Based on the analysis, we propose a training algorithm that
we call the Combined Search Algorithm (CS). It consists of two search
techniques and can be easily implemented in hardware.  Computer
simulations were conducted using IntraCardiac ElectroGrams (ICEGs) and
sonar reflection patterns and the results show that: using CS, the
training performance of feedforward multi--layer neural networks
implemented in digital hardware with 8 to 10 bit resolution can be as good as
that of networks implemented with unlimited precision; CS is insensitive to
training parameter variation; and importantly, the simulations confirm
that the numbers of quantization bits can be reduced in the upper layers
without affecting the performance of the network.
\end{abstract}


You can get the report by FTP:

unix>ftp 128.146.8.52
name:anonymous
Passord:neuron
ftp>binary
ftp>cd pbu/neuroprose
ftp>get yun.cs.ps.Z
ftp>bye
unix>uncompress yun.cs.ps.Z
unix>lpr yun.cs.ps


Yun