papers/neuroprose: Unifying MLP and RBFN

[Georg Dorffner]

FTP-host: archive.cis.ohio-state.edu
FTP-file: pub/neuroprose/dorffner.csfn.ps.Z pub/neuroprose/dorffner.nn-clinical.ps.Z

The files dorffner.csfn.ps.Z and dorffner.nn-clinical.ps.Z are now 
available for copying from the Neuroprose repository:

-------------------------------------------------------------------------
dorffner.csfn.ps.Z: 

A Unified Framework for MLPs and RBFNs:
Introducing Conic Section Function Networks

Georg Dorffner
Austrian Research Inst. for Artificial Intelligence
and
Dept. of Medical Cybernetics and AI
Univ. of Vienna
georg at ai.univie.ac.at


ABSTRACT: Multilayer Perceptrons (MLP, Werbos 1974, Rumelhart et al. 
1986) and Radial Basis Function Networks (RBFN, Broomhead & Lowe 1988, 
Moody & Darken 1989) probably are the most widely used neural network 
models for practical applications. While the former belong to a group 
of ``classical'' neural networks (whose weighted sums are loosely 
inspired by biology), the latter have risen only recently from an 
analogy to regression theory (Broomhead & Lowe 1988). On first sight, 
the two models -- except for being multilayer feedforward networks -- 
do not seem to have much in common. On second thought, however, 
MLPs and RBFNs share a variety of features, worthy of viewing them 
in the same context and comparing them to each other with respect to 
their properties. Consequently, a few attempts on arriving at a 
unified picture of a class of feedforward networks -- with MLPs and 
RBFNs as members -- have been undertaken (Robinson et al. 1988, 
Maruyama et al. 1992, Dorffner 1992, 1993). Most of these attempts 
have centered around the observation that the function of a neural 
network unit can be divided into a propagation rule (``net input'') 
and an activation or transfer function. The dot product (``weighted 
sum'') and the Euclidean distance are special cases of propagation 
rules, whereas the sigmoid and Gaussian function are examples for 
activation functions. This paper introduces a novel neural network 
model based on a more general conic section function as propagation 
rule, containing hyperplane (straight line) and hypersphere (circle)
as special cases, thus unifying the net inputs of MLPs and RBFNs with 
an easy-to-handle continuum in between. A new learning rule -- 
complementing the existing methods of gradient descent in weight 
space and initialization -- is introduced which enables the network 
to make a continuous decision between bounded and unbounded 
(infinite half-space) decision regions. The capabilities of CSFNs 
are illustrated with several examples and compared with exisiting 
approaches. CSFNs are viewed as a further step toward more efficient 
and optimal neural network solutions in practical applications.
                               
length: 37 pages.

submitted for publication

-------------------------------------------------------------------------
dorffner.nn-clinical.ps.Z:        

On Using Feedforward Neural Networks for Clinical Diagnostic Tasks

Georg Dorffner
Austrian Research Inst. for Artificial Intelligence
and
Dept. of Medical Cybernetics and AI
Univ. of Vienna
georg at ai.univie.ac.at

and

Georld Porenta
Dept. of Cardiology
Clinic for Internal Medicine II
University of Vienna

ABSTRACT: In this paper we present an extensive comparison between 
several feedforward neural network types in the context of a clinical 
diagnostic task, namely the detection of coronary artery disease 
(CAD) using planar thallium-201 dipyridamole stress-redistribution 
scintigrams. We introduce results from well-known (e.g. multilayer 
perceptrons or MLPs, and radial basis function networks or RBFNs) as 
well as novel neural network techniques (e.g. conic section 
function networks) which demonstrate promising new routes for future
applications of neural networks in medicine, and elsewhere. In particular 
we show that initializations of MLPs and conic section function networks
-- which can learn to behave more like an MLP or more like an RBFN -- 
can lead to much improved results in rather difficult diagnostic tasks.

Keywords: Feedforward neural networks, neural network 
initialization, multilayer perceptrons, radial basis function networks, 
conic section function networks; thallium scintigraphy, angiography, 
clinical diagnosis and decision making.

length: 21 pages

submitted for publication

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

To obtain a copy:

  ftp cheops.cis.ohio-state.edu
  login: anonymous
  password: <your email address>
  cd pub/neuroprose
  binary
  get dorffner.csfn.ps.Z
	AND/OR
  get dorffner.nn-clinical.ps.Z
  quit

Then at your system

  uncompress dorffner.*

to obtain (a) postscript file(s).

Many thanks to Jordan Pollack for the maintenance and support of this
archive.

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

both papers are also available through anonymous ftp from
ftp.ai.univie.ac.at in the directory 'papers' as
oefai-tr-93-23.ps.Z (== dorffner.nn-clinical) and oefai-tr-93-25.ps.Z
(== dorffner.csfn)

Hardcopies are available (only if you don't have access to ftp!) by 
sending email to sec at ai.univie.ac.at and asking for technical report
oefai-tr-93-23 or oefai-tr-93-25 (see previous paragraph).