\documentclass[12pt]{report}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{geometry}
\geometry{margin=1in}
\title{Integration and Hybridization in Neural Network Modelling}
\author{Wesley Royce Elsberry}
\date{August 1989}
\begin{document}
\begin{titlepage}
\centering
{\Large Integration and Hybridization in Neural Network Modelling\par}
\vspace{1.5cm}
{\large Wesley Royce Elsberry\par}
\vspace{1cm}
Presented to the Faculty of the Graduate School of\par
The University of Texas at Arlington in Partial Fulfillment\par
of the Requirements for the Degree of\par
\vspace{0.5cm}
{\large Master of Science in Computer Science\par}
\vfill
The University of Texas at Arlington\par
August 1989\par
\end{titlepage}
\chapter*{Acknowledgements}
I wish to thank the many people who have made my graduate program a

rewarding, enlightening, and interesting experience. My especial thanks

to Bob Weems, Ken Youngers, Vijay Raj, Steve Hufnagel, and Farhad Kamangar

for their exemplary instruction. The advice and encouragement of Bill

Buckles was critically important to this refugee from the life sciences.

The members of my graduate committee, Karan Briggs, Lynn Peterson, and

Daniel Levine, have provided me with technical resources, instruction,

referrals, and general good advice in plenty. Sam Leven provided the

example problem description, the classical sequence data, and much of his

own time and expertise to aid me in developing both the simulation program

and my understanding of the field. I am indebted to Dr. Levine for his

great interest in teaching the principles of cognitive modelling to as

wide an audience as possible, for without his express encouragement I

would not have become acquainted with the field, and also for his

considerable personal assistance in developing this thesis. The

enthusiasm of Harold Szu helped to motivate me to undertake a deeper

inquiry into neural network modelling. Finally, without the continual

support of my spouse, Diane Blackwood, this thesis and the classroom work

which formed the basis for it would not have been possible.

July 28, 1989

\chapter*{Abstract}
Artificial neural network models derived from different biological

behaviors or functions can be used in an integrative fashion to create an

extensive problem-solving environment. An example problem of limited

melodic composition is approached by the use of Hopfield-Tank, back-

propagation, and Adaptive Resonance Theory networks serving as plausible

next note generator, musical sequence critic, and novelty detector,

respectively. Biological bases for integrative function are discussed,

and the experimental role of synthetic systems such as the example

integrated network is explored.

\tableofcontents
\chapter{Defining The Role And Nature Of Artificial Neural Network Modelling}
Cultural Bias and Its Application to Cognitive Inquiry

Plato believed that there existed a pure world of ideas, whose

perfect forms were poorly mimicked by copies in the reality apparent to

our senses. This concept of the dominant, lofty nature of ideas and the

thoughts that manipulated those ideas would do more to inhibit inquiry

into material processes than virtually any other single cause in the

succeeding two millennia. The legacy of this philosophical outlook still

permeates our culture, coloring the basic assumptions and viewpoints of

researchers even after extended scientific training.

The firmness of general belief in the separate existence of ideas

contributed to the Lamarckian hypothesis of the inheritance of acquired

characters. While there were critics of Jean Baptiste Lamarck's work from

its introduction, these were by no means an overwhelming group by numbers.

The rediscovery of Gregor Mendel's work in 1905 provided clear and

convincing evidence against Lamarck's hypothesis, yet even Watson and

Crick's elucidation and characterization of DNA failed to dispel the last

vestiges of belief in the inheritance of acquired characters. The ready

and continued acceptance of the Lamarckian hypothesis into the mid-

twentieth century gives an indication of the continued influence of

Plato's world of ideas, even when confronted with contradicting evidence

and convincing counterarguments.

Biology, however, was by no means the only scientific discipline to

be touched by the cultural bias of Platonic ideals. Physics experienced

great upheaval and internal dissent as the deterministic view of Newton

and Laplace gave way to the quirkiness and Heisenbergian uncertainty of

quantum mechanics (Hawking 1988).

In thinking about thinking, researchers have tended to denigrate

approaches dependent upon investigation of physical processes. The high

regard given ideas and thought processes has nearly always provided

Western observers with a sense of revulsion at even considering that such

sacred items could be solely the product of soft, squishy brain tissues

and their component parts, neurons. This could be considered somewhat

akin to Roquentin's nausea upon consideration of radical contingency

(Ferguson 1987).

Our culture does not encourage modes of inquiry that tend to displace

the traditional view of the mind as the only known organ with which the

perfect world of ideas can be perceived. It especially does not encourage

the denial of the Platonic ideal world, yet it has been quite some time

since any reputable scientist has explicitly advanced support for that

concept.

Cognitive science, however, must deal explicitly and forthrightly

with the subject of ideas and thoughts. In a field where cultural bias

has, can, and will directly effect the discussion of the subject at hand,

it pays to recognize the existence and probable magnitude of that bias. I

have briefly delineated the persistence and magnitude of our Platonic bias

in biology and physics; I will assert that the bias is greater in

cognitive science, where it may be more directly challenged.

Artificial Intelligence: Definitions

Artificial intelligence describes at once a sub-discipline of

cognitive science and the goal of that sub-discipline. The goal is the

description and production of an artificial system or systems that can be

described as intelligent in operation, function, or effect, in both global

and local contexts. As with most complex fields, there are several ways

in which to approach direct inquiry into the topic. In what has been

termed the "top-down modelling school" of artificial intelligence, the

emphasis has been upon systems of formal logic and other explicit symbol

manipulation techniques. As may have been apparent from the previous

description, there is a "bottom-up modelling school" of artificial

intelligence. This school of inquiry seeks to examine the basic processes

that are known to produce intelligent action in humans, those basic

processes of neural function and coordination.

By extension, the bottom-up modelling school is concerned with neural

function in general. This is considered necessary because of the great

complexity of biological neural systems. While much research has been

done and continues to be done, there exists no basic understanding of the

detailed structure and operation of biological neural systems. There is a

wealth of data, but a paucity of organizing principles. The bottom-up

modelling school attempts to provide possible organizing principles,

testing these through a process of modelling and incremental design

improvements. This area of research has become known as artificial neural

network modelling.

The definitions given above differ somewhat from what may be

considered standard in the artificial intelligence community. In their

Turing Award speech, Newell and Simon state:

The notion of physical symbol system had taken essentially its present form by the middle of the 1950's, and one can date from that time the growth of artificial intelligence as a coherent subfield of computer science. The twenty years of work since then has seen a continuous accumulation of empirical evidence of two main varieties. The first addresses itself to the sufficiency of physical symbol systems for producing intelligence, attempting to construct and test specific systems that have such a capability. The second kind of evidence addresses itself to the necessity of having a physical symbol system wherever intelligence is exhibited. It starts with Man, the intelligent system best known to us, and attempts to discover whether his cognitive activity can be explained as the working of a physical symbol system. . . . The first is generally called artificial intelligence, the second, research in cognitive psychology. (Newell and Simon 1976)

This formulation of definitions seems at once too narrow and not properly

descriptive. It is too narrow in that it limits severely the range of

activities which may be considered artificial intelligence research.

While Newell and Simon talk of physical symbol systems, there is the

strong implication that they refer only to computational methods of

explicit symbol manipulation, as in the use of languages such as LISP and

Prolog. The definition of the second part is too narrow, in that it

postulates no overlap between the fields of artificial intelligence and

cognitive psychology, and also provides no linkage for concepts of

operation derived from empirical studies of biological intelligent systems

to be incorporated into an artificial intelligence framework. As stated

later by Newell and Simon,

The symbol system hypothesis implies that the symbolic behavior of man arises because he has the characteristics of a physical symbol system. Hence, the results of efforts to model human behavior

with symbol systems become an important part of the evidence for the hypothesis, and research in artificial intelligence goes on in close collaboration with research in information processing psychology, as it is usually called.

This seems to indicate that the previously given definitions were not

truly descriptive of the relationship between artificial intelligence and

other components of cognitive science. By including artificial

intelligence as a sub-discipline of cognitive science, it becomes clear

that artificial intelligence is not disjoint from research into

considerations of what constitutes intelligence.

An alternative formulation for defining artificial intelligence in

more general terms is given by Charniak and McDermott (1985), where they

state that artificial intelligence "is the study of mental faculties

through the use of computational models." This form of definition allows

for the top-down and bottom-up approaches to artificial intelligence to be

given co-equal rank as complementary research disciplines.

Artificial Neural Network Modelling

Artificial neural network modelling, the bottom-up school of

artificial intelligence, derives from the use of biological nervous

systems as suitable exemplars for an approach to coordination, cognition,

and control in artificial systems. This approach, in one form or another,

has been with us for a long time.

In the 1940's, McCulloch and Pitts demonstrated the possibility of

casting boolean logic systems into networks of thresholded logic units.

This certainly fits the Newell-Simon definition of a physical symbol

system. McCulloch and Pitts' seminal paper and subsequent work introduced

a new concept for consideration: rather than driving the understanding of

biological cognitive processes through increasing sophistication of

technology, perhaps technology can be made more capable and sophisticated

by elucidating mechanisms of function from biological processes involved

in cognition (Levine 1983, 1990).

This differs qualitatively from the viewpoint common to many

observers of human intelligence. There has been a tendency for people to

compare the operation of thought processes with the current leading edge

of technology. Freud compared the mind to a steam engine, Twain would

speak of "mill-works" in relation to speech production, and various

persons in this century have blithely and confidently settled upon the

computer as a fully analogous system. A circular system can be noted

here, though, as the image of the mind as being like some mechanism helped

contribute to such endeavors as Pascal's calculator, Babbage's Analytical

Engine, and the Hollerith punched card controlled census tabulator. These

efforts, in turn, inspired the modern computer. (The influence of the two

World Wars upon the motivation for the development of the computer must

also be acknowledged.) However, as the Metroplex Study Group on

Computational Neuroscience (1988) points out,

Computational study of brain structures began in the 1940's when digital computers were emerging. The first computer architectures were developed in a collaboration between the mathematicians John von Neuman and Norbert Wiener working closely with physiologists like Warren McCulloch. Early computer scientists were captivated by the analogy between the all-or-none action potentials of the neuron and binary switches representing bits in computers. Models of biological computation thus had a seminal role in the design of modern computer architecture.

Thus, while comparisons of the workings of the mind to technology helped

to motivate research into further technological advances, the insight into

the successful design for a significant new technology, computers, came

instead from paying attention to the actual mechanisms of brain function.

The computer has come to be pressed into service as a "new" metaphor

for thought processes in biological systems. While the use of

technological metaphors for cognition is persistent, having produced

quaint turns of speech and aphorisms, it has not been significantly

conducive to a direct understanding of the actual cognitive processes

being described. The turnabout analogy produced by McCulloch and Pitts

has provided far more benefits to cognitive science and computing through

its establishment of neural modelling. Binary switching technology failed

to provide significant insight into biological neural activity and

function. Leon Harmon (1970) noted that "[a]nother kind of difficulty in

coming to an understanding of nervous systems is that we may be

conditioned into thinking about them in ways that are more constrained

that we like to admit or are sufficiently aware of." The cultural

identity which we share through environment can be a handicap in research

into the nature and mechanisms of our intelligence.

The model networks of McCulloch and Pitts, an example of which is

shown in Figure 1, were seen to provide useful mechanisms and insights for

computing machinery. These networks consist of threshold logic units

receiving excitatory and inhibitory inputs of integer value. A simple sum

of these values is compared with a threshold level for the unit, and the

unit is considered to "fire" if its inputs equal or exceed the threshold

value. A unit's activity is then carried forward to further connected

units in the system. McCulloch and Pitts (1943) proved that any boolean

Figure 1, Mc-P

function could be created by the appropriate combination of such units.

The McCulloch-Pitts formalism is quite similar to principles of

design using TTL logic, as diagrams can attest. It should be noted that

McCulloch and Pitts described their formalism in 1943, some sixteen years

before the invention of the integrated circuit. While significant

problems reduce the overall utility of McCulloch-Pitts networks, modified

forms continue to be advanced (Szu 1989).

Given the basis of a neural computational framework, additional

features extracted from biological research were gradually added to

proposed models. The idea that inputs to neurons had graded values

modified by the efficiency of synaptic connections led to several

important advances. Hebb proposed a synaptic modification rule to allow a

form of learning in randomly connected networks (Hebb 1949, Levine 1983).

His formulation was that the efficiency of a synapse which connects two

neurons increases if both of the neurons have high activities at the same

time. This relatively simple rule has a number of drawbacks, which have

since been extensively cataloged and discussed. However, it cannot be

denied that Hebb's Law, as it has been called, was an important advance

for modelling.

The idea of stating an organizational principle for network design

also represented an important advance for artificial neural network

modelling. Rosenblatt's Perceptron architectures demonstrate this well.

Rosenblatt postulated several network models, among these was a three-

layer network. The layer which received external inputs was the sensory

layer, and the layer which gave an output (either as a raw signal or

interpreted as a motor output) was termed the response layer. An

associative layer of neurons provided the articulation between sensory and

response layers (Levine 1983). This general method of organization helped

move artificial neural network modelling away from specifically designed

networks and randomly organized network models.

The drawbacks associated with the special design of network models

remain a concern for ANN modellers today. Such systems are fragile or

brittle, meaning that a fault in any part of the network could cause a

general failure of function. Also, if an error in design occurred, the

resulting network would have no method of overcoming the design fault.

This particular pitfall is known by the assumption necessary for correct

function to be achieved: "programmer omniscience." Since programmer

omniscience cannot be guaranteed, systems predicated on this principle are

inherently unreliable (Hecht-Nielsen 1986).

Designing with generality as an important principle of function

results in artificial neural networks that are described as self-

organizing. A self-organizing network will conform its function to the

particular problem or context through a process of adaptation or learning.

Current Models Narrowly Focussed

The tendency in developing an artificial neural network model is to

constrain its function to a well-defined domain. This design principle

enables better control over evaluation of the functionality of the design.

Data from neurological or behavioral studies dealing with the problem

domain can then often be directly applied to either training or evaluating

the model.

There exist many general functions that are commonly encountered in

biological neural networks which can be considered to have important

implications for computational study. Among these are included

associative memory, classification, pattern recognition, and function

mapping. ANN models proposed to implement these functions include Bi-

directional Associative Memory, Adaptive Resonance Theory

(classification), Brain-State-in-a-Box (classification), Neocognitron

(invariant pattern recognition), and Back-Propagation (function mapping)

(Simpson 1988).

Unfortunately, the underlying reason for existence of each model is

overlooked in comparing different models. The human tendency to wish to

attribute a single scalar quantity denoting how "good" an ANN model is

will often cause people to overlook the fact that certain networks should

not be compared as being equivalent. Lippmann provides a good basic

overview of six different ANN models, yet falls prey to this pitfall.

Each of the models is interpreted as a classification network, yet only

the Hamming and ART networks are designed to function as classifiers

(Lippmann 1987).

Multi-architecture Integrative Systems

The use of multiple complementary architectures in designing

application systems has not often been explored. While this approach to

system design is commonly touted in ANN simulation aid advertisements, it

is less frequently featured in the literature. In an advertisement for

the ANZA Neurocomputer, HNC Corporation states, "In these networks the

interconnect geometry is already determined and the form of the transfer

equations is fixed. However, the number of processing elements (neurons),

their initial state and weight values, learning rates, and time constants

are all user selectable, thus allowing one to customize a particular

network paradigm (or combination of paradigms) to suit a particular

application." There are some reasons for the relative lack of multi-

paradigm (multi-architecture) research results in publication.

The design principles stated earlier still hold, that it is easiest

to design an architecture with a narrow definition of function. There is

the tendency for such architectures to incorporate simplifying assumptions

which aid in simulation or real-world implementation ("casting in

silicon," as the phrase goes). The ease of verifying correct operation is

increased for architectures which have a narrowly defined application, as

relatively clean data or simple theoretical findings are likely to be

available against which the architecture may be evaluated.

The relative complexity of modelling for even constrained contexts

provides another reason for concentration upon single paradigm systems.

The number of parameters which can effect a model's performance ranges

from zero for linear systems, such as Widrow's ADALINE (Widrow 1987), to

tens of scalar parameters, as may be encountered in Carpenter and

Grossberg's Adaptive Resonance Theory architecture (Carpenter and

Grossberg 1987a, 1987b; Simpson 1988). When dealing with nonlinear

dynamical systems for which no closed-form solution exists, slight changes

of system parameters can result in large scale changes in behavior.

Carpenter and Grossberg point out relative constraints upon the parameters

used in the ART 1 architecture, giving guidelines for values which should

yield stable operation of the network. Finding suitable parameters for

operation of a particular architecture can be a frustrating and time-

intensive experience.

As an example, an architecture called the "on-center, off-surround"

network (OCOS) is based upon a relatively simple equation that appears in

slightly modified form in many competitive networks. One form of this

equation (Grossberg 1973) is:

dx/dt = -A Xi [first term]

+ (B-Xi) (Ii+f(Xi)) [second term]

- (Xi+C) (Sum from k = 1 to n of ((k<>i) (Jk+f(Xk))) [third term]

(Eq. 1)

The first term represents decay of activity over time; the second

term represents increase in activity due to excitatory input values, Ii,

and recurrent self-excitation; the third term represents decrease in

activity due to inhibitory input values, Jk and competition from other

nodes in the network. For each term, there is an associated parameter. A

hidden parameter of the architecture is a factor by which to multiply the

result of the difference equation. This is usually set to be much less

than one, especially when using the simple forward Euler method of system

updating. However, there is a performance trade-off involved with setting

this factor too small: the system will converge to a stable result, but

will take a long time to converge. If the factor is set to a large value

the network will "blow-up," which is a state of wild fluctuation in

activation values for the nodes in the network. Such a system cannot

converge. The solution is to find a value for this factor which will

ensure convergence without an inordinate amount of time taken in reaching

convergence. The explicit parameters given in the equation must be

matched to the size of the other parameters to maintain set relationships.

In order to find parameter values, I have used a network simulation which

runs through permutations of sets of discrete parameter values. By

examining the resulting output which indicated whether the network run

achieved convergence and the number of time steps necessary to achieve

convergence, I was able to settle upon a suitable choice of parameters for

regular use and was able to decipher a coarse set of relationships between

the parameters.

In consideration of these complexities in modelling and simulation,

synthesis of complex systems from simpler subunits has some significant

obstacles to overcome. There are more degrees of freedom for system

operation, leading to complexities in articulation and coordination of

subunits. This may increase the necessary number of system parameters,

increasing complexity in a combinatorial fashion. Usually, there will

either be less data available for evaluation of the complete system

behavior, or the data that is available will be uncertain. The general

principles to apply to a synthetic endeavor in artificial neural network

modelling remain to be elucidated.

In addition, synthesis has traditionally been given short shrift in

Western culture (Paris 1989). The application of synthesis in philosophy

has largely been confined to the works of Wittgenstein and Marx. Marx, of

course, developed a philosophy whose application was inimical to most

Western economic structures. This has caused both his work and his

approach to be considered with contempt.

On the positive side, however, synthesis can lead to significant

insights as different assumptions and features can be applied to

functional design. The process of integration and hybridization in

artificial neural network modelling may be expected to lead to new

insights into structure and function of benefit to artificially

intelligent systems, given the past history of benefits derived from

consideration of concepts borrowed from disparate disciplines. In this

case, the complexity of knowledge or topic of a contributing discipline

can provide a measure of the possible beneficial overlap: as complexity

increases, the set of concepts which may profitably be applied to the new

context also increases. As this rule would predict, biological nervous

systems provide a very large pool of concepts ready for reconstruction in

the artificial intelligence framework.

Biological Neural System Complexity

The complexity which drives that expectation can be readily

appreciated, as in the human nervous system the number of neurons is in

the billions, and each neuron will have connection to between tens and

tens of thousands of other neurons. This physio-spatial complexity pales

when compared to the complexity that can result when one attempts to

define a state for the neuron. The state can be considered to be a

combination of electrical activity, ionic balance, hormonal levels, input

activities, and many other factors. Since for several of these attributes

there exist separate contributing factors, the number of components

contributing to neuron state can easily exceed twenty, and there may be

hundreds of such components. Add to this that these components often have

analog values, separate time scales of action, and capacity for permanent

change in the neuron's response, and the possible state function for a

single neuron displays a suitably bewildering complexity.

Since this level of complexity is not conducive to direct

examination, the process of problem decomposition and solution is applied.

The implication for artificial neural network modelling is that certain

gross features are modelled. Typically, a model assumes that the major

interesting component of neural function is the electrical activity of the

contributing neurons or nodes. A further assumption in general currency

is that the electrical activity of an ANN may be modeled with the elements

or nodes responding in the manner expected of neural populations rather

than individually thresholded neurons (Levine 1983). Some models take

into account generalized neurotransmitter effects, with Grossberg's gated

dipole model providing a good example (Grossberg 1972). These simplifying

assumptions still leave a high level of complexity to be dealt with in

developing ANN designs and applications.

Relevance of Biological Models

Biology provides the exemplars for self-organizing adaptive systems

which have proven useful in artificial neural network modelling thus far.

While strict adherence to biological accuracy in modelling may well be

counter-productive to advances in modelling techniques, rejecting

biological principles may then lead to difficulties in later integrative

work.

The biological framework of neurophysiological function provides a

basic structure which makes for a common ground of interaction in models.

For example, the range of activation and output of neurons is rather

strictly proscribed, while the range of activation of an analogous unit in

neural network models is limited only by the capability of the processor

in specifying a floating point exponent. This would make interfacing two

models using different ranges of activation less straightforward.

The inherent complexity of biological systems can support two

different arguments concerning the future of modelling. As Hawking (1988)

says, describing a complex system all at once is terrifically difficult.

It is much simpler to model several different components of the overall

system. Bringing these partial models together to form a complete model

can become problematic, as in Hawking's example of the search for the

Grand Unified Theory in physics. In one sense, then, the effort to

integrate models of limited function may be premature. However, it may be

that along with the incremental advances made in modelling small subunits

of biological neural function, we should also attempt incremental

integration of well-understood low level models. This would help prevent

the kind of situation which exists in physics, with two or more highly

disjoint models prevailing and no unification or integration yet in sight.

Since biological systems are inherently more complex than the physical

systems upon which they are based, it becomes important to keep an eye on

the eventual need for integration and resolution of subsidiary models.

Models which utilize features from two or more extant ANN

architectures such as Hecht-Nielsen's Counterpropagation Network (Hecht-

Nielsen 1987) demonstrate the useful qualities which synthesis of models

can bring to functionality. The counterpropagation network is derived

from Kohonen's self-organizing map model and Grossberg's competitive

learning networks. Hecht-Nielsen notes that this network is designed for

function mapping and analyzes its performance as compared to the back-

propagation architecture. While the general functionality of counter-

propagation networks (CPN) remains lower than that of back-propagation

networks, there exists a subclass of mapping functions for which the CPN

will train faster, and there also exists a closed-form solution for the

error of the CPN. As Hecht-Nielsen notes, "Finally, CPN illustrates a key

point about neurocomputing system design. Namely, that many of the

existing network paradigms can be viewed as building block components that

can be assembled into new configurations that offer different information

processing capabilities."

Artificial Neural Network Models: Three Architectures

For the purposes of exploring multi-architecture system design, I

selected three artificial neural network architectures for incorporation

into the overall system. These were the Hopfield-Tank network, back-

propagation, and Adaptive Resonance Theory 1 models. These networks

perform three different functions: Hopfield-Tank is used for optimization

or constraint satisfaction, back-propagation is a general-purpose mapping

network, and Adaptive Resonance Theory 1 is a classifier network.

Hopfield-Tank Networks

In 1982, John Hopfield wrote an article, "Neural networks and

physical systems with emergent collective computational abilities," which

was published in Proceedings of the National Academy of Sciences,

describing a model of neural computation which was readily implementable

in current solid-state technology. This article and the model which it

described has been widely credited with a resurgence of interest in ANN

modelling.

The network architecture presented by Hopfield and Tank (1985),

hereafter known as HTN for "Hopfield-Tank Network[s]", is a single-layer

fully interconnected network model (Figure 2). There is no learning rule

for this network, although various researchers have proposed modified HTN

architectures that do incorporate adaptive learning. Weights between

nodes in an HTN are fixed and symmetric, and connections between a node

and itself are zero. The advantages of these criteria for network design

are that the system dynamics can be shown to perform an "energy

minimization" in reaching a stable state. When the weights are determined

according to system constraints, the system can be characterized by a

Liapunov function, providing a measure for system energy.

HTN's have been applied to various constraint satisfaction and

optimization problems. Hopfield and Tank attracted much attention by

demonstrating the utility of an HTN in generating good solutions to the

"Traveling Salesman Problem" (or TSP), a non-polynomial time complete

problem. The TSP can be described as choosing a minimum length path among

Figure 2

a set of cities such that each city is visited once, and the salesman

returns to the city of origin. The closed path length constitutes the

measure for any particular solution. As the number of cities increases,

the number of possible valid tours increases combinatorially. However, it

can be shown that an HTN will produce good tours in constant time. An HTN

used for TSP computation does not necessarily converge upon the optimum

solution, but will reject non-optimum solutions and give relatively "good"

solutions. Unfortunately, it can also be shown that the HTN trades off

constant time operation for O(n2) space considerations.

Figure 2 shows the HTN architecture for solution of the TSP.

Although the nodes in the network are treated as elements in a vector, the

visual representation which makes the most sense to the designer and other

interested humans is a matrix of nodes. By imposing this structure upon

our view of the network, we can associate each row with a city in the

tour, and each column with the position of a city in the final tour (City

n is in Position m in the tour). The constraints of importance are that

for a valid tour, each city appears once, and each position in the tour

has only one city. Thus, for the state of the network at equilibrium, we

should find high activities in n nodes, where n is equal to the number of

cities or positions. The interconnections between nodes in a row or

column are inhibitory, causing highly active nodes to reduce the activity

of other nodes in the same row or column. The interconnections from a

node to its neighbors in adjacent columns are proportionally more

inhibitory as distance between cities represented by those nodes

increases.

Hopfield-Tank Equation

The equation defining the network's activity over time (Hopfield and

Tank 1985) is

Ci (dui/dt) = (Sum from j = 1 to N of (TijVj - ui/Ri + Ii)) (Eq. 2)

where i and j designate neurons in the network,

N is the total number of neurons in the network,

T is a connection weight between neurons,

V is the output value function for a neuron,

u is the activity of a neuron,

I is the external input for a neuron,

C is the capacitance of a neuron,

and R is the resistance of a neuron.

Back-propagation Networks

Back-propagation is a short version of "correction by the backward

propagation of errors." The learning rule used in back-propagation

networks (BPNs) is termed the generalized delta rule (Simpson 1988;

Rumelhart, Hinton, and Williams 1986). Basically, a BPN is a multi-layer

(with at least three layers) network whose nodes use a sigmoid output

function. The BPN will map a set of input activities to another set of

output activities, given training upon a set of example input/expected

output vector pairs. The BPN will generally discriminate and adapt to

non-linear relationships in the training data. For example, a BPN can

learn the exclusive-or relation, which a single-layer perceptron cannot.

Nodes in the BPN are often called "units."

The basic premise of the back-propagation algorithm has been

independently derived several times. Werbos (1974) is generally credited

with the first publication of the learning rule, which he called dynamic

feedback; Parker (1985) gave the rule the name, "learning logic;" and the

popularity of the BPN architecture is primarily attributable to Rumelhart,

Hinton, and Williams (1986), as related by Simpson (1988).

All units in the BPN operate in basically the same manner. There are

some slight differences dependent upon whether a unit resides in the

input, hidden, or output layer. Generally, however, a unit generates an

output signal as follows:

oj = f(netj) (Eq. 3)

where j is a unit in the BPN,

net represents the input to a unit,

f is a sigmoid function,

and o is the output of a unit.

netj = (Sum over i of (oiwij)) (Eq. 4)

where i is a unit in a preceding layer of the BPN,

and w is a connection weight linking two units.

f(netj) = 1 / (1 + e(-(netj + thetaj))) (Eq. 5)

where theta is the bias weight for the unit.

An input unit has a net input which is simply the provided external input,

as there are no preceding layers.

Figure 3 demonstrates a three layer BP network. At the bottom are

five input units, which receive their activation from an external source.

In the middle are sixteen hidden units, which receive their input

according to Equation 4. The hidden units' output is sent on over another

set of weights to the single output node.

The error between the external training signal and the output node's

output activity provides the basis for correcting the behavior of the

network. This raw measure of error is used to find the delta for the

output node.

deltak = (tk - ok) f'(netk) (Eq. 6)

where k is an output unit,

t is the expected output,

f' is the derivative of the output function.

Figure 3

Fortunately, the derivative of the sigmoid function is symbolically easy

to specify:

f' = f (1 - f) (Eq. 7)

The network then must distribute this error measure backward through the

network. For each of the hidden units, then, deltas are found as follows:

deltaj = (Sum over k of (deltakwjk)) f'(netj) (Eq. 8)

where k is a unit in the succeeding layer of the BPN.

Each of the weights in the network is changed according to:

(Change in wij) = L * oideltaj (Eq. 9)

where L is a constant representing the learning rate for the BPN.

Theta values for nodes are treated in the same manner, thus

(Change in thetaj) = L * f(thetaj)deltaj (Eq. 10)

So, for the example BPN in Figure 3, an input vector causes output of

the input units to be distributed to the hidden units, modified by the

intervening weights. Similarly, the hidden units send on their output

through weights to the output unit, which provides the response of the

network to the particular input vector. This is called "feed-forward"

processing.

Once the network output is known, it can be compared to the expected

output, and the delta value for the output unit is determined. This

begins the "back-propagation" process. The delta values for the hidden

units can now be derived as in Equation 8. The amount of change for each

of the weights between the hidden and output layers can now be found

according to Equation 9. Weight changes between input and hidden layers

proceed similarly. For each node in the hidden and output layers, the

value of theta is also changed. This completes the normal back-

propagation phase.

In the example problem, we have made a slight change to the normal

back-propagation process, and have allowed theta for the input units to

also be adaptively changed.

In practice, one would normally construct a network with the correct

numbers of input and output units, make some guess as to the number of

hidden units needed, and assign random values to the weights. The network

would then be trained upon the available data vector pairs until the error

becomes suitably low, or the implementor decides to make a change in

network design.

Possible applications for BPNs include encrypting, data compression,

non-linear pattern matching and feature detection. Existing BP

applications include translation of text inputs into phoneme outputs,

acoustic signal classification, character recognition, speech analysis,

motor learning, image processing, knowledge representation, combinatorial

optimization, natural language, forecasting and prediction, and multi-

target tracking. BP has been implemented or theorized in electronic,

VLSI, and optical formats (Simpson 1988).

Adaptive Resonance Theory 1

Adaptive Resonance Theory 1 (ART 1) is a model introduced by Gail

Carpenter and Stephen Grossberg (Carpenter and Grossberg 1987a). There

are two ART models of note, ART 1 and ART 2, and many modified

architectures which are premised upon one or the other of the ART models.

Basically, an ART architecture is a two-layer network which provides

unsupervised learning of categories of inputs (Figure 4). The F1 layer is

composed of "feature nodes," which accept external inputs. In ART 1,

these inputs are binary patterns, while ART 2 incorporates preprocessing

to accept analog inputs to the F1 layer. The F2 layer is composed of

"category nodes," which compete to respond to valid F1 activations. There

are control structures built into the architecture to prevent F2

activation without input being received at the F1 layer. There are other

control structures to prevent "resonance" from occurring when the

prototype pattern determined by the most active F2 node does not

correspond to the pattern of F1 activation.

Short term and long term memory are represented in ART architectures

by node activations and inter-node weights, respectively. A "bottom-up

activation" refers to the pattern of activation received by F2 nodes

through the weighted links from F1 nodes. Similarly, a "top-down

activation" is the pattern of activation received by F1 nodes through

weighted links from F2 category nodes. Long term memory is changed only

through resonance between the F1 nodes and a selected F2 node. A more

Figure 4

detailed diagram of the ART 1 interconnections is shown in Figure 5.

Resonance is a state in which long term memory traces between the F1

and F2 layers are modified to more closely represent the input activation

in the category node's top down weights. An F2 node which wins the

competition among F2 nodes has its top-down activation tested against the

bottom-up activation of the F1 feature nodes. If these match within a

level of tolerance, called the vigilance level, a resonant state is

entered and long term memory is changed. If not, the current winning F2

category node is made ineligible for further consideration against the

input, and F2 competition is restarted among remaining eligible F2 nodes.

For an example pattern presented to the ART network given in Figure

4, the presence of input turns on the gain control nodes and activates the

F1 layer. The activation of the F1 layer, or bottom-up activation, is fed

across a set of long term memory weights to generate a new pattern of

activity at the F2 layer. The F2 layer responds with a top down

activation which is filtered by the top down weights linking each of the

F2 nodes to the F1 layer. Competition among the F2 nodes results in a

single winner for application to the current feature input. A comparison

of the bottom up activation with the top down activation yields a set

theoretic measure of the match between the presented pattern and the

category represented by the F2 node. If I represents the input pattern,

V(J) represents the F2 category node J's archetypal pattern, and p

represents the vigilance parameter, then the cardinality of the set

intersection between pattern I and the category pattern V(J) must be

greater than or equal to p times the cardinality of I, or reset occurs.

If reset occurs, the F2 node J becomes ineligible for matching to the

current input pattern, and the process of bottom up activation, top down

activation, competition, and match testing continues until some category

node is found to match the input sufficiently well, or until all category

nodes have been matched against the input and found to be too different.

Figure 5 displays a more detailed ART 1 network, with the various weights,

nodes, and connections visible.

Obviously, it is possible that none of the eligible F2 nodes will

match within the vigilance level's acceptable tolerance. When an ART

network is first trained, there are no eligible F2 nodes, rather there are

a number of uncommitted F2 nodes. An F2 node will be selected and enter a

resonant state, providing the first category. If a subsequent input does

not match this category within the vigilance level, the single F2 category

is rendered ineligible and the F2 layer is reset. This brings the network

to a state analogous to that of having no eligible category nodes

available, and a second category node is selected and resonated with the

F1 layer. At any point where no further eligible F2 nodes exist, but an

uncommitted F2 node remains, then a new category node is formed from the

formerly uncommitted node. If no uncommitted category nodes remain, then

the input has been found not to match the available categories.

Figure 5

\chapter{Integration In Neural Network Modelling}
Integration in neural network modelling is taken here to mean the

combination of different neural network architectures in a coordinated

system. This differs from the casual usage sometimes found in the

literature where the term has been applied either to systems composed of

multiple units of the same base architecture, or else to trivial

modifications of a known architecture. Integration applies properly to

cases of multiple-architecture systems and there are some instances of

systems for which the term genuinely applies but has not been used.

As an example, Matsuoka, Hamada, and Nakatsu (1989) have proposed an

architecture for phoneme recognition that subdivides the hidden and output

layers of a back-propagation network in order to enhance the network's

ability to recognize phonemes and also to substantially reduce the

training time necessary for the network. However, Matsuoka terms this

architecture the Integrated Neural Network (INN). While there is a

substantial improvement in training time, there is no fundamental

difference between an INN and a back-propagation network: they differ only

in the connectivity of the weights between the hidden and output layers.

The reduction in internal complexity of the INN can explain the decreased

training time discovered.

A slightly different form of integration is pursued by Foo and Szu

(1989). A "divide and conquer" approach to problem solving employs the

same architecture, a modified Hopfield-Tank network, to handle smaller

subproblems and then brings together the resulting subproblem solutions

into an overall solution. This requires some coordination to effect the

overall solution, bringing into play elements of the broader integrative

issues I have noted, but is not properly an integrated system as I have

defined it.

A better example of an integrated artificial neural network system

appears in Cruz et al. (1989). Cruz uses a MADALINE architecture for

image preprocessing and a back-propagation network for removing image

distortion. The MADALINE architecture is a Multiple ADALINE system,

developed by Widrow (Widrow, Pierce, and Angell 1961). The ADALINE, short

for Adaptive Linear Neuron, is a neuronal model using the Least Mean

Square learning rule developed by Widrow and Hoff. Widrow and Winter

(1988) have updated the learning rule used for multiple layer, multiple

ADALINE networks. The specific example given by Widrow and Winter

presents a MADALINE network for invariant pattern recognition. The

MADALINE architecture most closely resembles the Perceptron architecture

given by Rosenblatt. In creating his integrated system, Cruz applied the

"divide and conquer" concept not for the purpose of reducing simulation

time, but in consideration of space requirements needed for a system which

could handle the 256x256 pixel images used.

Integration and Convergence

A perennial problem for the artificial neural network modeller is the

issue of convergence in finite time. It is nice to know that the

architecture selected for a function will converge to a solution before

the heat death of the universe. It is similarly a concern that a system

composed of subsidiary architectures will converge. This can be

problematic, since general convergence theorems have not been found for

several of the most popular architectures (Widrow 1987).

Back-propagation provides a particularly pointed example, since it is

by far the most popularly used network, if number of papers concerning

applications is taken as the criterion. The generalized delta learning

rule of back-propagation has long been appreciated to be generally useful,

yet significant progress in firmly establishing this usefulness in the

form of theorems concerning convergence has been lacking. Sontag and

Sussman (1989) provide a theorem demonstrating for back-propagation an

analogous result to the perceptron learning rule: if a separation solution

exists, the generalized delta rule's gradient descent will find the

solution in finite time. While much has been made of a theorem by

Kolmogorov (Farhat 1986), it must be conceded that Kolmogorov's theorem is

an existence theorem: there is some network based upon the back-

propagation architecture which will perform a mapping from n to m, but the

theorem gives no clues as to what that specific network looks like.

There is some promising news concerning convergence which is of

interest for building integrative ANN systems. Hirsch (1989) gives

several theorems which hold that if component subnets of a neural network

converge, then the network will converge. While Hirsch's theorems assume

that the convergence properties of the subnetworks can be described by

Liapunov energy functions, he notes, "It is more difficult to obtain

convergence for cascades of systems that are merely assumed to be

convergent, but without benefit of Liapunov functions or global asymptotic

stability. One way of doing this is to place strong restrictions on the

rates of convergence." Hirsch defines a cascade as a layered network

where the output of one layer serves as the input of the next. Many

integrative designs can be cast into this framework.

Incremental Synthesis

The process of synthesis leading to integrative artificial neural

network modelling is important to the development of insights into topics

of critical application, such as sensor fusion. By confronting directly

the need for coherent internal use of available resources and

capabilities, we are more likely to generate an understanding of fusion

principles.

The synthetic approach to modelling provides a supportive environment

for creating extensive systems. Just as the topic of artificial neural

network modelling benefits from the interdisciplinary nature of its

supporting sciences, so a synthetically derived artificial neural network

system benefits from the range of problem solution approaches and features

inherent in the underlying network architectures. By creating and

maintaining a system of network architectures applicable to subfunctions

in the problem solution, subfunction solution by a particular system

component can benefit from the co-option of features normally found in

different components of the artificial neural network system.

Networks Under Consideration, System Properties

Each of the three network architectures used in the example problem

of melodic composition has its own set of features and drawbacks which

play an important role in system design.

Hopfield-Tank Networks

As mentioned earlier, the Hopfield-Tank architecture is generally

used in achieving a specific function. By this I mean that each Hopfield-

Tank network is designed for a particular purpose, and can provide no

functionality for other unrelated purposes. So the use of a Hopfield-Tank

network should generally be reserved for functions which do not change

over time. However, I will immediately cite a counter-example, due to its

elegance of integrative design.

An ingenious mechanism for extending the utility of the Hopfield-Tank

architecture is pursued by Tsutsumi (1989), where one back-propagation net

remaps Hopfield-Tank network inputs and another back-propagation network

remaps Hopfield-Tank network outputs. The problem given is one of

avoiding robotic arm deadlocking. The movement space is constrained, and

therefore applicable to Hopfield-Tank network solution. However, the

adaptation of the internal space representation to the real world arm

movement must be adaptive. The Hopfield-Tank network does not provide

learning rules, so the back-propagation networks provide the adaptation to

real world feedback. In this manner, some significant benefits accrue to

the use of the integrated system: since the Hopfield-Tank network is

static with respect to the encoded weights, it provides a good repository

for the robot's joint-arm space; since the back-propagation networks are

adaptive, the system can configure itself to respond to a changing

environment.

Since the Hopfield-Tank network was conceived of in the context of

implementation in silicon, the possibility for reducing a Hopfield-Tank

network instance to a hardware component can be important for real-world

applications. This step would bring the benefit often touted for

Hopfield-Tank networks, speed. Speed is rarely noted in practice, since

most practice involves simulation. The simplicity of the Hopfield-Tank

architecture can be a strong point for system design and integration,

however, even in simulated systems.

A drawback which may eliminate the Hopfield-Tank network from

consideration for a particular function is that the weights for the

network must be derived from the constraints to be implemented and from

any data functions necessary for solution but not available in the input

to the network. This requires an understanding of the system to be

solved, which may not be available. Some architecture with learning rule

would then be more suitable for application.

The output of a Hopfield-Tank network must be deciphered from the

final pattern of activation of the net at equilibrium (cf. discussion in

Chapter 1). In the case of the Travelling Salesman Problem, position of

active nodes provides an encoding of placement of cities in the tour.

This information might be rendered more compact in another format,

depending upon the input type expected for further networks in the system.

Back-propagation Network System Considerations

The back-propagation architecture provides a mapping from the input

vector to the output vector, and can be trained by example. The system

properties of back-propagation networks include stability of learning

given a fixed universe. By implication, the learning is not stable if

some perturbation in the problem set changes the mapping function. The

back-propagation network would then learn the new mapping over time, and

the old mapping would be lost.

Back-propagation networks have a moderately complex structure. The

properties gained from this increase in complexity over the Hopfield-Tank

architecture include the capacity for learning, the ability to extract

features from input data and generate internal representations for those

features, and the possibility of complex input to output transforms in

accordance with learned associations.

Back-propagation networks can accept binary or analog inputs, so the

inputs can represent conditional probabilities as well as more strictly

constrained values. The outputs are analog values which can be

interpreted as binary through the use of thresholding functions. This

allows a wide variety of input and output possibilities to achieve overall

system function. However, the choice of representation of inputs can be

critical for speedy and reliable training to occur. For example, use of

analog values should be avoided when there exists a natural partition of

the range of the input into distinct states. Our example of the melodic

note generator will illustrate this concern.

ART 1 System Properties

Adaptive Resonance Theory 1 architectures provide unsupervised

learning of "clusters" or classifications of input vectors. An internal

representation of classification archetypes is generated. This

architecture ensures that new inputs will tend not to perturb

classifications of previous inputs. This compromise in the stability-

plasticity tradeoff (Carpenter and Grossberg 1987a, b) can be modified for

special purposes, as the melodic note generator program will demonstrate.

The internal operation of the ART 1 network can provide certain features

of especial interest to system design. Specifically, one by-product of

the classification algorithm is the detection of novelty, which will be

shown to have functional significance beyond that of the original design

of the ART 1 network.

Some properties of ART 1 require particular attention from the

designer. For example, there is a matching parameter (vigilance) which

controls how much deviation from a category prototype is acceptable.

There are no guidelines for the selection of the vigilance parameter, and

it is left to the designer to select and assign a "proper" value. Some

guidelines do exist for certain of the other learning parameters in the

ART 1 architecture, such as the learning coefficients for each of the top-

down and bottom-up memory equations (Carpenter and Grossberg 1987a).

The ART 1 architecture provides no standard output. The designer

must access internal values of the ART 1 network to provide useful

information to the remainder of the system which includes that network.

ART 1 is a highly complex architecture with many parameters to be

selected and set by the designer. Useful modifications to be made to the

architecture would include creating adaptive functions to replace some of

the static and arbitrary parameters of the network, such as the gain and

reset parameters (cf. Figure 5).

Table 1. System Properties Summary

Hopfield-Tank network properties: - Data initialization process only place for changing adaptively (not "learning" at all) - Fast convergence (on system, not necessarily on simulation) - Inflexible structure (individual design necessary) - Simple structure

Back-propagation Network Properties: - Stable learning given fixed universe - Change to design implies relearning necessary (costs for self-adaptation include forgetting what has been learned) - Adaptive weights changed as consequence of training (supervised learning) - Medium complexity of structure - Input to output transform can be computationally complex - Output nodes may deliver digital, bipolar, or analog values

ART Network Properties: - Stable self-organizing learning - Change in design produces unknown effect on existing "knowledge" - Non-adaptive parameters for gain and reset possible drawbacks, should be replaced by adaptive functions

System considerations

Considerations of interfacing outputs of one model to inputs of another: Inputs: HTN : analog or discrete value, one per node BPN : analog or discrete value, one per input node ART 1: discrete value, one per input node ART 2: analog value, one per input node

Outputs: HTN : discrete value, one per node BPN : analog or discrete value, one per output node ART 1 : none specified ART 2 : none specified

\chapter{Example Problem}
In developing a simulation to test out processes of integration,

several factors have to be considered. The simulation should have enough

scope to provide good subsidiary roles for the component functions, it

should be small enough to be implementable in a reasonable period of time,

and it should produce output of a form which is readily apprehensible and

analyzable. The first criterion is easily fulfilled; the second and third

are rather more difficult to fulfill.

As a starting point for exploring the possibilities of an integrated

approach to artificial neural network modelling, the problem of producing

a melodic line in music composition was selected. The complexity of music

composition in general provides ample considerations for the application

of component networks. Unfortunately, the complexity of musical

composition offers no hope for the relatively simple design and

implementation of an artificial neural network system which addresses all

the salient points. Therefore, simplifying assumptions are made to ease

the requirements for the ANN system. The output, to be interpreted as a

sequence of musical notes, can present some problems with evaluation, due

to the qualitative context of musical evaluation in general. However, it

is possible to treat the note sequence as a set of concatenated symbols,

and apply some of the information theory concepts of Shannon (1948) to

conduct an analysis.

Simplifying Assumptions

The great complexity of musical composition in general is constrained

to yield a problem of suitable scope for the example integrated ANN

system. A limited scale covering one octave is assumed, in the key of C.

There are no accidentals, and there is no explicit timing of notes. A

single voice is assumed, and there are no harmonics generated. A limited

and fixed set of classical composition rules forms the basis for the

constraint and comparison of the system.

Problem Approach

The use of several ANN architectures in creating an integrated system

whose function fulfills the requirements of the musical composition

problem is assumed. A preliminary set of hypotheses as to how a composer

develops a line of melody was advanced for defining the subfunctions of

the composition system, or note generator. As Teuvo Kohonen (1989) notes

in discussing his own ANN system for musical composition,

It is not possible to survey here the development of ideas in computer music. One of the traditional approaches, however, may be mentioned. It is based on Markov processes. Each note (pitch, duration) is thereby regarded as a stochastic state in a succession of states. The probability Pr = Pr(Si | Si-1, Si-2,

% raw formatter directive: ...) for state Si, on the condition that the previous states are
Si-1, Si-2, ..., is recorded from given examples. Usually three predecessor states are enough. New music is generated by starting with a key sequence to which, on the basis of Pr and, say, the three last notes, the most likely successor state is appended. The augmented sequence is used as a new key sequence, and so the process generates melodic successions ad infinitum. Auxiliary operations or rules are necessary to make typical forms (structures) of music out of pieces of melodic passages.

Leaving the problem of constructing musical forms for possible later

consideration, the approach given by Kohonen matches well the procedure of

composition undertaken here. A candidate note and note sequence are

proposed, a critique according to classical rules is made concerning the

last note in the candidate sequence. The approval of the critic should

mostly cause the acceptance of the candidate note, but the rules should be

broken often enough that there is not an absolute conformity to the rules

postulated.

Now we must confront the problem of how to best combine models in a

unified structure that accomplishes the example function of melodic

composition. The solution involves identifying the strengths and

weaknesses of the component models, and which functions each may

accomplish. Then when the subproblem mapping has been accomplished, it

must be determined how to integrate the subfunction module outputs to

other modules to accomplish the overall task.

Example Problem Network System

The identifiable problem subfunctions are candidate note generation,

sequence critique, and novelty detection. The candidate note generator

subfunction should propose notes in general, but not complete, conformance

to probabilities of the next note filling the requirement of being part of

a classical sequence. The sequence critique subfunction should evaluate

the proposed next note in strict conformance to the set of classical

sequences provided. Since the evaluation given by the sequence critique

subfunction is an incomplete criterion for composition given expectations

of novelty, some means of detecting novel sequences must exist for the use

of the coordinating system. The coordinating system then has the

information necessary to "break the rules" when needed to avoid long,

boring sequences of strictly mechanical melody. This can also be seen as

a requirement for stable and continued operation, as there exist sequences

which have no classical next note possibilities.

Candidate Note Generation

The candidate note generator should produce plausible next notes

given a historical partial sequence. Since the rules for identification

of plausible next notes are fixed and known, there is no need for learning

in this stage of the network. The candidate note generator should also

occasionally provide notes which do not necessarily conform to the

expectation of a classical next note.

The Hopfield-Tank network (HTN) was selected as the candidate note

generator since it provides the above features. HTNs are noted for their

utility in constraint satisfaction and other optimization problems, which

fits the requirement that only a single next note is to be proposed at a

given time and that it should basically follow the probabilities of a

classical next note. A well known attribute of the HTN architecture is

the inclusion of spurious local minima which do not represent "valid"

solution states. By purposefully utilizing this feature, we can convert

what normally constitutes a drawback into a asset for the subfunction.

The spurious local minima will give the occasional proposal of non-

classical next notes. The constant and known nature of expected sequences

determines the formation of the HTN weights, in conjunction with the known

constraints for proper HTN function. The use of "noisy" input values can

produce the semi-random distribution of possible notes that is needed for

variability.

Figure 6 shows the modified HTN architecture, called Bach, used in

our simulation. The rows represent note values and the columns represent

sequence placement. The constraints imposed on this network include the

need to present a single winning note in each place in the sequence

pattern, and the need to prevent endless repetition of the same candidate

note. By introducing relatively strong inhibitory links within rows and

columns, we can satisfy the constraint requirements. We achieve

preferential selection of classical next notes by reducing those

inhibitory links somewhat for connections which follow classical sequence

patterns.

Sequence Critique

The sequence critic should provide an evaluation of the conformance

of the proposed next note to classical melodic sequence rules. If the

sequence rules are assumed to be provided by example, then some learning

function is required to allow the critic to become adept and reliable.

The subfunction receives a note sequence as input, and produces an output

which may be interpreted as a boolean statement of the classical nature of

the candidate note.

The back-propagation network (BPN) was selected as the sequence

Figure 6

critic. The BPN, as discussed in Chapter 2, can learn any input/output

transformation given to it (within some constraints upon the availability

of sufficient hidden nodes to form a stable internal representation).

The form of the BPN used, called Salieri, was in the end a network

accepting a binary representation of the input sequence, using twenty

hidden units for internal representation, and producing an output value

which was interpreted as a yes-or-no output. So the net structure used

forty input units, twenty hidden units, and one output unit. This is a

medium sized BPN.

Novelty Detection

The novelty detection function must have components to enable the

recollection of past sequences for discrimination of novel sequences.

However, the space requirements should not be overwhelming. A

classification system would provide good data compression while allowing

the nearly complete context information needed for novelty detection.

The ART 1 architecture was selected for the role of the novelty

detector, for it provides novelty detection as part of a classification

framework. The ART 1 architecture also has other features of interest to

further research on integrated and extensive systems.

The ART 1 network used here, called Beethoven, is modified from the

Carpenter-Grossberg architecture. Some of the explicit features of the

Carpenter-Grossberg architecture are handled as implicit assumptions in

the procedural simulation. The "2/3rds Rule," for example, is not

invoked, since the only time that Beethoven is active, the network meets

the constraints of the rule. The separate rules for top-down and bottom-

up weight modifications are replaced by a single rule for both, as is done

in the ART 2 architecture (Carpenter and Grossberg 1987b).

Coordination

In any integrative system, some means of coordinating subfunctions

becomes necessary. Some interpretation and processing of input or output

terms may be accomplished by the coordinating system. The ultimate

decision for whether or not a candidate note is accepted falls to the

coordinating system.

The coordinating system is called Lobes, since the features and

activities of the frontal lobes (Levine 1986, Levine and Prueitt 1989)

provided the inspiration for its operation. Lobes generates the context

management and state dependent actions which drive the integrated system

to completion of the intended function, melodic composition.

Lobes also contains an internal boredom function, which tends to

increase over time. There is a boredom threshold, which causes a change

in behavior in Lobes if it is exceeded.

Operation

In operation, the integrated ANN note generator uses its components

in a sequential manner. Lobes generates a call to Bach for a candidate

note, which when returned is sent to Salieri for critique. As a mechanism

for preventing wastage of Beethoven category nodes, the entire sequence

which Bach settles upon is used for further processing by Salieri and

Beethoven. Since at the beginning of composition there is no sequence

history, Bach's inputs are determined randomly over the entire sequence.

As notes are added to the sequence history, Bach inputs are determined

with the addition of some noise, except for the candidate note column

which receives only random noise as input.

Salieri receives the Bach sequence values and makes a determination

of classical conformance, which it passes back to Lobes. Lobes then sends

on the entire context, sequence plus critique, to Beethoven. At first,

Beethoven is virtually certain to encode new categories for each input it

receives. As time goes by, the likelihood that Beethoven will encode a

new category decreases, until all category nodes are utilized. So it is

more likely that the first few notes generated will conform to classical

sequence examples. Sometimes, however, a note which has no possible

classical successors will be generated early in the simulation. In this

case, it is likely that Beethoven's category encoding will proceed at a

much faster than average pace.

The indication of novelty generated by Beethoven is used by Lobes to

modify the system response to the internal state, which is determined by

Salieri's critique of the next note, Beethoven's detection of novelty, and

the boredom threshold in Lobes. If Lobes is not bored and Salieri

approves of the candidate note, the note is accepted. If, however, Lobes

has reached its boredom threshold and Salieri approves of a note, it will

reject the candidate note and request another one from Bach. Likewise, if

Salieri disapproves of the note but Lobes is bored, Salieri may be

overridden and Lobes may accept the note. Indications of novelty from

Beethoven can satisfy Lobes' drive to be "excited," or not-bored. This

will tend to make Lobes more conservatively classical as Beethoven

continues to detect novelty. Figure 7 shows the operation of the

coordinated system.

Figure 7

\chapter{Results}
Performance

The integrated ANN note generator system produces 152 notes in about

three hours when run on an 80386 PC compatible at 16MHz clock speed. It

takes approximately fifteen hours to produce the same number of notes on

an 8088 based machine with an 8087 numeric coprocessor.

Example and Analysis of Output

Appendix A contains sample output from a run of the note generator.

With a problem such as musical composition, assigning an objective measure

denoting the "worth" of the output is not possible. However, it is

possible to compare the output of our note generator system with random

sequences of note values. By use of a binomial performance measure, it is

possible to define how much the sample output differs from a random

sequence.

Random sequences have their own mystique and interest, but subjective

evaluations of random melodic forms by the untrained ear tend toward the

negative. The output of our note generator network was intended to

basically follow the guidelines of an example set of classical sequences.

The system included mechanisms for breaking out of a strict adherence to

the guide set of sequences. Thus, it would be expected that the output

have somewhat more resemblance to random sequences than the "rules" would

state. Since the classical rules of composition, while not our sole

criteria of fitness, are the only quantifiable part of our criteria, we

compare our output with random note sequences on the basis of these rules.

Table 2 summarizes the characteristics of output sequences generated

under several different conditions. A random note generator, a mostly

classical note generator, and our integrated ANN note generator each

produced outputs and were evaluated using the critic developed for

training Salieri. The "Successes" column indicates the number of times

the next note in the sequence could be considered to be classical.

The random note generator simply output a sequence of random numbers

for a set sequence length. The Classical Instructor sequence generator

operated by determining the available pool of possible classical next

notes, then randomly selecting one of those notes. In some cases, no next

note fit the criteria of being "classical," and a random note was

generated. The rest of the sequence generators were variations upon our

ANN note generator. The use of different back-propagation nets in the

critic role gave different results in the output. Trained and untrained

back-propagation nets with analog inputs were used. No significant

difference in output could be distinguished between the trained and

untrained versions, but the result was far closer to random performance

than classical. The inability of the Salieri net with analog inputs to

converge to a reliable and accurate performance measure explains the

similarity of result with the untrained version of the same type. On the

other hand, a Salieri composed of a trained back-propagation network using

a binary representation for inputs was able to converge to a stable and

fairly accurate state. Hence the performance of the trained Salieri with

binary inputs was considerably closer to classical than was the

performance of the random sequence generator. In two cases, a rule-based

critique system was substituted for the back-propagation network.

Table 2. Classical components of output sequences.

Sequence generator Sequence Successes \% Z versus Z versus Length random classical

Random 10,000 890 0.089 0.00 -102.54

System w/ untrained Salieri (analog inputs) 152 18 0.118 1.26 -23.84

System w/ trained Salieri (analog inputs) 152 19 0.125 1.54 -23.64

System w/ trained Salieri (binary inputs) 152 36 0.237 6.28 -20.07

System w/ rule-based critique (Salieri supervisor) Run 1 150 35 0.233 6.10 -20.06 Run 2 150 47 0.313 9.42 -17.50

Classical Instructor 8,150 6898 0.846 102.54 0.00 (rule-based critique w/out ANN system)

\chapter{Discussion}
The integrated note generation system's performance suggests that it

met operational expectations. It produced notes according to a mixture of

somewhat conflicting criteria. This kind of operation in the midst of

uncertainty characterizes many human decision-making processes, and may be

assumed to play a role in human music composition as well. Our framework

allows for further experimentation with hypotheses concerning the

fundamental processes involved in higher-order constraint satisfaction

systems in an extensive environment (cf. Pao 1989).

The integrated approach has demonstrated several advantages and

disadvantages in development and operation. The disadvantages include the

complexity of handling several different network architectures at once,

which can contribute to programmer confusion (the downfall of programmer

omniscience). The necessity for dealing directly with "model mismatches,"

where one subnetwork may produce a different representation as output than

the next subnetwork requires as input, can cause system design time to be

protracted. Failure to recognize that a problem exists in articulation of

networks can result in behavior that diverges wildly from expected norms.

On the other hand, any complex problem may present similar

difficulties regardless of the choice of solution approach. By using

subnetworks of known characteristics, one may be able to achieve a

solution with fewer uncertainties than a totally top-down approach would

yield. With a range of different capabilities available from subsidiary

networks, the likelihood of encountering an insoluble subproblem is

reduced. Synthetic integrated systems also lead to a combinatorial

explosion in the richness of possible system behaviors, which again is

reminiscent of the increasingly interesting behaviors noted as more

complex biological organisms are considered.

Simulation Concerns

The integrated ANN system from the example problem relied on several

procedural programming shortcuts. For example, the implementation of

"boredom" in Lobes was simply a counter which would be compared with a

threshold value. This does not have any basis in biological neural

systems, yet rather neatly simulates the behavior of a simple neuronal

model for the same task. As another example, the decision in the ART 1

network as to which category an input belongs to is currently made on the

basis of an arbitrary, winner-take-all rule. The same effect could

probably be achieved by means of on-center-off-surround neural

interactions that are more biologically realistic. Casting the system

functions into an entirely biological framework would yield a better, more

capable system for future work. However, the system stands as a first

effort toward this goal.

Integration and Artificial Intelligence

There have been several efforts toward integrative system design in

the top-down modelling school. In the HEARSAY system (Reddy et. al.

1973), the blackboard model was developed. The blackboard model

presupposes the combination of a set of possible problem-solving systems

which have available the current system state. The system state is said

to reside upon "the blackboard" as a visualization of the process of

problem-solving. Each of the several applicable subsidiary problem-

solving systems attempts to derive an incremental step toward the global

solution, and competes with other such problem-solvers to control the

blackboard, and thus be able to change the system state. This is a

significant development, and one which is paralleled by concepts in

several artificial neural network models. In ART models, for example, the

concept of competition among various classification prototypes bears a

strong resemblance to the blackboard model. If the F1 layer's activation

is considered analogous to the system state, the F2 nodes each are

analogous to problem-solvers in the blackboard model.

There have been some efforts toward explicitly bringing together top-

down and bottom-up models in hybrid systems. Amano et al. (1989) present

an example of a phoneme recognition system combining an expert system with

a perceptron network. The expert system provides feature extraction from

speech data, which is input to the perceptron. The perceptron allows

decision-making under uncertainty, whose output is interpreted using fuzzy

logic rules. This avoids drawbacks associated with "template matching"

phoneme recognition schemes. Rabelo and Alptekin (1989) have integrated a

neural network with an expert system into an intelligent scheduling system

for manufacturing applications. Their system has the ability to learn

from experience and generate schedules within real-time constraints.

Neurobiological Evidence of Multifunctionality

Multi-state functionality of memory is supported by work of Nottebohm

(1989). Nottebohm's work involves the development and remembrance of

complex songs in songbirds. Through a series of studies, Nottebohm

demonstrates that hormonal changes can cause the forgetting of songs in a

bird's repertoire, or allow the formation and remembrance of new songs.

The hormonal changes in question center around testosterone, and typically

the mating season is the time when levels of testosterone allow the

formation of songs, presumably conferring a reproductive advantage for

male songbirds. Nottebohm demonstrates that the changes are only

hormonally dependent by the simple strategy of artificially inducing song

creation by the application of testosterone to songbirds of both sexes at

various times of the year. The withholding of regular doses of

testosterone was also shown to cause the forgetting of known songs in the

same birds.

The implications of Nottebohm's work include support for state-

dependent memory. Since a specific memory function can be modulated by a

specific hormone, this implies that other memory systems may also have

recall dependent upon some hormone or other chemically mediated process.

Given that recall and learning can be so modulated, the necessity for

taking state dependencies (levels of hormones and other neuroactive

substances) into account for system function can be appreciated.

State-dependent memory is also supported by the work of many other

investigators, such as Bower (1981). In Bower's experiments, happy or sad

moods were induced in subjects by hypnotic suggestion, in order to

investigate the influence of emotions on memory and thinking. This

influence was profound; for example, people recalled a greater percentage

of those experiences which had taken place when they were in the same mood

as they were in during recall. Also, when the feeling tone of a story

agreed with a reader's hypnotically generated emotion, the reader found

the events and characters in the story more memorable and easier to

identify with.

State dependence of memory or other neural function can give rise to

quite useful modelling constructs. The ability to recast problem

solutions given functional states becomes biologically justifiable. The

logical power of conditional activation of entire subnetworks becomes

available through the modelling, however coarse, of these state

dependencies. The almost direct implementation of expert system analogues

which can be analyzed in a completely biological and ANN context is made

possible.

Implications include the higher order integration of functionally

changed sub-units over time. In humans, well known state dependencies

include the fight or flight response to norepinephrine production and the

diving response typical of mammals entering cold water. Without the

appropriate integrative control, neither fight or flight nor the diving

response would produce the desired, or selected, effect. The coordination

of separate functional neural "circuits" is clearly present; the exact

mechanism remains to be elucidated but there have been some promising

beginnings in neural network models. For example, in the neural model of

attention described by Grossberg (1975), there is competition between

nodes representing activations of different drives (hunger, thirst, sex,

etc.). The winner of this competition is not determined solely by which

drive is highest, but also by the availability of compatible cues in the

environment.

State-dependent memory implies the existence of functional changes

over time in cortical structures. Since we now have evidence of multi-

modal neural circuitry, at least some consideration should be given to the

implied necessary integration. The problem of understanding a system

which is dynamic not only in processing of input but also in functional

neural subsystems is both daunting and exhilarating. It is daunting,

because the complexities of modelling such systems exceeds our current

capacity for ready assimilation and understanding of the underlying

concepts and mechanisms (which have not yet been elucidated), and

exhilarating because there appears to be no end to the variety of

expression of these systems in the natural world, and thus no apparent end

to the problem-solving challenge awaiting the researcher.

The function of speech processing in humans, for example, requires

the acquisition of external signals, the separation of those signals into

semantic and affective content, the recognition of mode in affective

content, the parsing of semantic content, and the integration of semantic

and affective content to determine meaning. This list of subfunctions is

not complete, which gives an indication of the extent to which integration

remains a regular and important activity in biological systems.

The Triune Brain Theory

Integrative theories of neural/cognitive function have a long

history. One of the best known is the triune brain theory of Paul MacLean

(1970). MacLean's research into behavioral studies of different brain

areas led him to propose that the human brain is divided into three

developmentally derived regions of separate function (see Levine 1990 for

discussion). The earliest, and presumably the most primitive region, is

termed the reptilian brain, and is composed of the the brainstem and basal

ganglia. The reptilian brain is responsible, in this theory, for the

preprogrammed, innate behaviors. The paleomammalian brain, composed of

the limbic system, modifies the expressed pattern of reptilian brain

responses and is the source, in this theory, of the basic emotions (love,

hate, fear, arousal, etc.). The neomammalian brain, composed of the newer

parts of the cerebral cortex, provides further modifications of the

expression of the two older brain areas, and gives us our rational

capacity, seen in the ability for planning and verbal expression.

While MacLean's theory is oversimplified, it does provide a useful

set of distinctions between various cognitive subfunctions, all of which

are involved in complex behaviors. In fact, if one stretches the

imagination, one can draw analogies between the reptilian brain and our

Hopfield-Tank network; the paleomammalian brain and our ART 1 network, and

the neomammalian brain and our back-propagation network.

The integrated ANN note generator had its origins in a collaborative

effort to develop an extensive ANN system suitable for exploring multi-

modal cognitive hypotheses (Blackwood, Elsberry, and Leven 1988). That

project, in turn, was derived from insights provided by Leven (1987b).

Leven's SAM model was depicted in a manner which led to a discussion of

the possibility of replacing the components of MacLean's triune brain

model with current ANN architectures. The difficulty of describing a

suitably restricted problem for adequate application of the limited

current architectures was resolved with the simple melodic composition

problem outlined previously. Points of difference from the original

MacLean theory can be attributed to certain changes in model context (as

modified by Leven's separation of memory into three components:

motoric/instinctive [reptilian], sensory/affective [paleomammalian], and

associative/semantic [neomammalian] (Leven 1987a)) and to the mismatch

between architectures derived not for their similarity to these basic

cognitive forms, but to satisfy more immediate criteria such as being

implementable in current electronic devices. The desired system based on

our loose analogy to MacLean's theory has been demonstrated to be

operational and ready for incremental refinement.

We hope that in future work such analogies can be made more precise.

The development of both neural network theory and neuroscientific data

should allow the critical research to continue into these theories of

integrative cognitive function.

\appendix
\chapter{Sample Melody Output Of The Various Note Generator Programs}
Integrated ANN Note Generator Sample Output

b61T output, page 1

b61t output, page 2

b61t output, page 3

Random Note Generator Sample Output

random output, page 1

random output, page 2

random output, page 3

Classical Note Generator Sample Output

classical output, page 1

classical output, page 2

classical output, page 3

\chapter{Program Source Listing: Integrated Ann Note Generator}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Back-Propagation Unit}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: List Structures Unit}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Miscellaneous Procedures Unit}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Global Type And Variables Unit}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Classical Instructor Unit}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Ansi Screen Control Unit Unit Ansi\_Z;}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Musical Sequence Evaluator Program}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Random Note Generation Program}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Note Sequence Playing Program}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Rule-Based Note Sequence Generation Program}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Offline Back-Propagation Network Training Program}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Data File Listing: Hopfield-Tank Network Weight Data File}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Data File Listing: Classical Sequences Data File}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Data File Listing: Back-Propagation Network Data File}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\chapter{Program Source Listing: Translator From Program Note Files To Music Transcription System Song Format}
This appendix is represented in the repository by the legacy source and data files in \texttt{THES/}. The automated thesis conversion suppresses the full listing here to keep the document manageable.

\nocite{*}
\bibliographystyle{plain}
\bibliography{integration_and_hybridization_in_neural_network_modelling}
\end{document}