HM Medical Clinic

 

To_editor.dvi

Computer Science Department IBM Almaden Research Center Michigan State University East Lansing, MI 48824 San Jose, CA 95120 Numerous advances have been made in developing intelligent" programs, some of which have been inspired by biological neural networks. Researchers from variousscientic disciplines are designing articial neural networks (ANNs) to solve a varietyof problems in decision making, optimization, prediction, and control. Articial neuralnetworks can be viewed as parallel and distributed processing systems which consistof a huge number of simple and massively connected processors. There has been aresurgence of interest in the eld of ANNs in recent years. This article intends to serveas a tutorial for those readers with little or no knowledge about ANNs to enable themto understand the remaining articles of this special issue. We discuss the motivationsbehind developing ANNs, main issues of network architecture and learning process, andbasic network models. We also briey describe one of the most successful applicationsof ANNs, namely automatic character recognition.
1 IntroductionWhat are articial neural networks (ANNs)? Why is there so much excitement about ANNs? What are the basic models used in designing ANNs? What tasks can ANNs perform eciently? These are the main questions addressed in this tutorial article.
Let us rst consider the following classes of challenging problems of interest to computer scientists and engineers.
Pattern classication: The task of pattern classication is to assign an input pattern (e.g., speech waveform or handwritten symbol) represented by a feature vector to one of pre- specied classes (Fig. 1(a)). Well-known applications of pattern classication are character recognition, speech recognition, EEG waveform classication, blood cell classication, and printed circuit board inspection.
Clustering/categorization: In clustering, also known as unsupervised pattern classi- cation, there are no training data with known class labels. A clustering algorithm explores the similarity between the patterns and places similar patterns in a cluster (see Fig. 1(b)).
Well-known clustering applications include data mining, data compression, and exploratory data analysis.
Function approximation: Given a set of n labeled training patterns (input-output pairs), (x ;y );(x ;y ); ;(xn;yn) , generated from an unknown function (x) (subject to noise), the task of function approximation is to nd an estimate, say , of the unknown function  (Fig. 1(c)). Various engineering and scientic modeling problems require function Prediction/forecasting: Given a set of n samples y(t );y(t ); ;y(tn) in a time sequence, t ;t ; ;tn, the task is to predict the sample y(tn ) at some future time tn .
Prediction/forecasting has a signicant impact on decision making in business, science and engineering. Stock market prediction and weather forecasting are typical applications of prediction/forecasting techniques (see Fig. 1(d)).
Optimization: A wide variety of problems in mathematics, statistics, engineering, sci- ence, medicine, and economics can be posed as optimization problems. The goal of an optimization algorithm is to nd a solution satisfying a set of constraints such that an objec- tive function is maximized or minimized. A classical optimization problem is the Traveling Salesperson Problem (TSP), which is an NP-complete problem.
Content-addressable memory: In the Von Neumann model of computation, an entry in memory is accessed only through its address which is independent of the content in the memory. Moreover, if a small error is made in calculating the address, a completely dierent item would be retrieved. Associative memory or content-addressable memory, as the name implies, can be accessed by its content. The content in the memory can be recalled even by a partial input or distorted content (see Fig. 1(f)). Associative memory is extremely desirable in building multimedia information databases.
Control: Consider a dynamic system dened by a tuple u(t);y(t) , where u(t) is the control input and y(t) is the resulting output of the system at time t. In model-referenceadaptive control, the goal is to generate a control input u(t) such that the system follows a desired trajectory determined by the reference model. An example of model reference adaptive control is the engine idle speed control (Fig. 1(g)).
A large number of approaches have been proposed for solving the problems described above. While successful applications of these approaches can be found in certain well- constrained environments, none of them is exible enough to perform well outside the domain for which it is designed. The eld of articial neural networks has provided alternative ap- proaches for solving these problems. It has been established [1, 8, 6] that a large number of applications can benet from the use of ANNs.
Articial neural networks, which are also referred to as neural computation, network computation, connectionist models, and parallel distributed processing (PDP), are massively parallel computing systems consisting of an extremely large number of simple processors with many interconnections between them.
The purpose of this article is to serve as a tutorial for those readers with little or no knowledge about articial neural networks. The rest of this article is organized as follows.
Section 2 provides the motivations behind developing ANNs. In Section 3, we describe the basic neuron model, network architecture, and learning process. Sections 4 through 7 pro- vide more details about several well-known ANN models: multilayer feedforward networks, Kohonen's self-organizing maps, ART models and the Hopeld network. In Section 8, we discuss character recognition, a popular and one of the most successful applications of ANN models. Concluding remarks are presented in Section 9.
Pattern Classifier over−fitting to noisy training data Airplane partially Retrieved airplane occluded by clouds Figure 1: Tasks that neural networks can perform. (a) Pattern classication; (b) clustering/categorization; (c) Function approximation; (d) Prediction/forecasting; (e) Optimization (TSP problem); (f) Retrieval by content; and (g) Engine idle speed 2 MotivationANNs are inspired by biological neural networks. This section provides a brief introduction to biological neural networks.
2.1 Biological Neural NetworksA neuron (or nerve cell) is a special biological cell with information processing ability. A Figure 2: A sketch of a biological neuron.
schematic drawing of a neuron is shown in Fig. 2. A neuron is composed of a cell body, or soma, and two types of out-reaching tree-like branches: axon and dendrites. The cell body has a nucleus which contains information on hereditary traits and a plasma containing molecular equipment for the production of material needed by the neuron. A neuron receives signals (impulses) from other neurons through its dendrites (receivers), and transmits signals generated by its cell body along the axon (transmitter) which eventually branches into strands and substrands. At the terminals of these strands are the synapses. A synapse is a place of contact between two neurons (an axon strand of one neuron and a dendrite of another neuron). When the impulse reaches the synapse's terminal, certain chemicals, called neurotransmitters are released. The neurotransmitters diuse across the synaptic gap, and their eect is to either enhance or inhibit, depending on the type of the synapse, the receptor neuron's own tendency to emit electrical impulses. The eectiveness of a synapse can be adjusted by the signals passing through it so that the synapses can learn from the activities in which they participate. This dependence on past history acts as a memory which is possibly responsible for the human ability to remember.
The cerebral cortex in humans is a large at sheet of neurons about 2 to 3 mm thick with a surface area of about 2,200 cm , about twice the area of a standard computer keyboard.
The cerebral cortex contains about 10 neurons, which is approximately the number of stars in the Milky Way! Neurons are massively connected, much more complex and denser than today's telephone networks. Each neuron is connected to 10 10 other neurons. In total, the human brain contains approximately 10 Neurons communicate by a very short train of pulses, typically milliseconds in duration.
The message is modulated on the frequency with which the pulses are transmitted. The frequency can vary from a few up to several hundred Hertz, which is a million times slower than the fastest switching speed in electronic circuits. However, complex perceptual deci- sions, such as face recognition, are made by a human very quickly, typically within a few hundred milliseconds. These decisions are made by a network of neurons whose operational speed is only a few milliseconds. This implies that computations involved cannot take more than about one hundred serial stages. In other words, the brain runs parallel programs that are about 100 steps long for such perceptual tasks. This is known as the hundred step rule [5]. The same timing considerations show that the amount of information sent from one neuron to another must be very small (a few bits). This implies that critical information is not transmitted directly, but captured and distributed in the interconnections, and hence the name connectionist model.
Interested readers can nd more introductory and easily comprehensible material on biological neurons and neural networks in [3].
2.2 Why Articial Neural Networks?Modern digital computers have outperformed humans in the domain of numeric computation and related symbol manipulation. However, humans can eortlessly solve complexperceptual problems (e.g., recognizing a person in a crowd from a mere glimpse of his face) at such a high speed and extent as to dwarf the world's fastest computer. Why does there exist such a remarkable dierence in their performance? The biological computer employs a completely dierent architecture than the Von Neumann architecture (see Table 1). It is this dierence that signicantly aects the type of functions each computational model is best able to Von Neumann computer Biological computer separate from a processor integrated into processor non-content addressable content addressable numerical and symbolic perceptual problems Table 1: Von Neumann computer versus biological computer.
Numerous eorts have been made to develop intelligent" programs based on Von Neu- mann's centralized architecture. However, such eorts have not resulted in any general- purpose intelligent programs. ANN models are inspired by biological evidence, and attempt to make use of some of the organizational" principles that are believed to be used in the human brain. Our ability to model a biological nervous system using ANNs can increase our understanding of biological functions. The state-of-the-art in computer hardware technology (e.g., VLSI and optical) has made such modeling feasible.
The long course of evolution has resulted in the human brain possessing many desirable characteristics which are neither present in a Von Neumann computer nor in modern paral- lel computers. These characteristics include massive parallelism, distributed representation and computation, learning ability, generalization ability, adaptivity, inherent contextual in- formation processing, fault tolerance, and low energy consumption. It is hoped that ANNs, motivated from biological neural networks, would possess some of these desirable character- The eld of articial neural networks is an interdisciplinary area of research. A thorough study of articial neural networks requires a knowledge about neurophysiology, cognitive sci- ence/psychology, physics (statistical mechanics), control theory, computer science, articial intelligence, statistics/mathematics, pattern recognition, computer vision, parallel process- ing, and hardware (digital/analog/VLSI/optical). New developments in these disciplines continuously nourish the eld of ANNs. On the other hand, articial neural networks also provide an impetus to these disciplines in the form of new tools and representations. This symbiosis is necessary for the vitality of neural network research. Communications among these disciplines ought to be encouraged.
2.3 Brief Historical ReviewResearch in ANNs has experienced three consecutive cycles of enthusiasm and skepticism.
The rst peak, dating back to the 1940's, is due to McCullough and Pitt's pioneering work [14]. The second period of intense activity occurred in the 1960's which featured Rosenblatt's perceptron convergence theorem [18] and Minsky and Papert's work showing the limitations of a simple perceptron [16]. Minsky and Papert's results dampened the enthusiasm of most researchers, especially those in the computer science community. As a result, there was a lull in the neural network research for almost 20 years. Since the early 1980's, ANNs have received considerable renewed interest. The major developments behind this resurgence include Hopeld's energy approach [9] in 1982, and the backpropagation learning algorithm for multilayer perceptrons (multilayer feedforward networks) which was rst proposed by Werbos [20], reinvented several times, and popularized by Rumelhart et al. [19] in 1986.
Anderson and Rosenfeld [2] provide a detailed historical account of developments in ANNs.
3 Articial Neural NetworksThis section provides an overview of ANNs. First, computational models of neurons are introduced. Then, the important issues of network architecture and learning are discussed.
Various ANN models are organized by their architecture and the learning algorithm involved.
3.1 Computational Models of NeuronsMcCulloch and Pitts [14] proposed a binary threshold unit as a computational model for a neuron. A schematic diagram of a McCulloch-Pitts neuron is shown in Fig. 3. This Figure 3: McCulloch-Pitts model of a neuron.
mathematical neuron computes a weighted sum of its n input signals, xj; j = 1;2; ;n, and generates an output of 1" if this sum is above a certain threshold u, and an output of 0" otherwise. Mathematically, where ( ) is a unit step function at zero, and wj is the synapse weight associated with the jth input. For simplicity of notation, we often consider the threshold u as another weight w = u which is attached to the neuron with a constant input, x = 1. Positive weights correspond to excitatory synapses, while negative weights model inhibitory synapses.
McCulloch and Pitts proved that with suitably chosen weights a synchronous arrangement of such neurons is, in principle, capable of universal computation. There is a crude analogy here to a biological neuron: wires and interconnections model axons and dendrites, connection weights represent synapses, and the threshold function approximates the activity in soma.
The model of McCulloch and Pitts contains a number of simplifying assumptions, which do not reect the true behavior of biological neurons.
The McCulloch-Pitts neuron has been generalized in many ways. An obvious generaliza- tion is to use activation functions other than the threshold function, e.g., a piecewise linear,sigmoid, or Gaussian, shown in Fig. 4. The sigmoid function is by far the most frequently used function in ANNs. It is a strictly increasing function that exhibits smoothness and has the desired asymptotic properties. The standard sigmoid function is the logistic function, g(x) = 1=(1 + exp( x) ; where  is the slope parameter.
Figure 4: Dierent types of activation functions.
3.2 Network ArchitectureAn assembly of articial neurons is called an articial neural network. ANNs can be viewed as weighted directed graphs in which nodes are articial neurons and directed edges (with weights) are connections from the outputs of neurons to the inputs of neurons. Based on the connection pattern (architecture), ANNs can be grouped into two major categories as shown in Fig. 5: (i) feedforward networks in which no loop exists in the graph, and (ii) feedback (or recurrent) networks in which loops exist because of feedback connections. The most common family of feedforward networks is a layered network in which neurons are organized into layers with connections strictly in one direction from one layer to another. Fig. 5 also shows typical networks of each category. We will discuss in this article all these networks except for the Radial Basis Function (RBF) networks [6] which employ the same network architecture as multilayer perceptrons, but dierent activation functions.
Dierent connectivities yield dierent network behaviors. Generally speaking, feedfor- ward networks are static networks, i.e., given an input, they produce only one set of output Feedforward Networks Figure 5: A taxonomy of network architectures.
values, not a sequence of values. Feedforward networks are memoryless in the sense that the response of a feedforward network to an input is independent of the previous state of the network. Recurrent networks are dynamic systems. Upon presenting a new input pattern, the outputs of the neurons are computed. Because of the feedback paths, the inputs to each neuron are then modied, which leads the network to enter a new state.
Dierent network architectures require dierent learning algorithms. The next section will provide a general overview of the various learning processes.
3.3 LearningThe ability to learn is a fundamental trait of intelligence. Although a precise denition of learning is often dicult to state, a learning process in the context of articial neural networks can be viewed as the problem of updating network architecture and connection weights so that a network can eciently perform a specic task. Most of the time, the network must learn the connection weights from the available training patterns. Improvement in performance is achieved over time through iteratively updating the weights in the network.
The ability of articial neural networks to automatically learn from examples makes them very attractive and exciting. Instead of having to specify a set of rules, ANNs appear to learn them from the given collection of representative examples. This is one of the major advantages of neural networks over traditional expert systems.
In order to understand or design a learning process, one must rst have a model of the environment in which a neural network operates, i.e., what information is available to the neural network. We refer to this model as a learning paradigm [6]. Second, one must understand how weights in the network are updated, i.e., what are the learning rules which govern the updating process. A learning algorithm refers to a procedure in which learning rules are used for adjusting weights in the network.
There are three main learning paradigms, namely, (i) supervised, (ii) unsupervised, and (iii) hybrid learning. In supervised learning or learning with a teacher, the network is provided with a correct answer to every input pattern. Weights are determinedso that the network can produce answers as close as possible to the known correct answers. Reinforcement learning is a variant of supervised learning where the network is provided with only a critique on the correctness of network outputs, not the correct answers (outputs) themselves. In contrast,unsupervised learning or learning without a teacher does not require any correct answer associated with each input pattern in the training data set. It explores the underlying structure in the data, or correlations between patterns in the data, and organizes patterns into categories from these correlations. Hybrid learning combines supervised learning and unsupervised learning. Typically, a portion of weights in the network are determined using supervised learning, while the others are obtained from unsupervised learning.
Learning theory must address three fundamental and practical issues associated with learning from samples: (i) capacity, (ii) sample complexity, and (iii) time complexity. Capac-ity concerns how many patterns can be stored, and what functions and decision boundaries can be formed by a network.
Sample complexity determines the number of training patterns needed to train the net- work in order to guarantee a valid generalization. Too few patterns may cause over-tting" (wherein the network performs well on the training data set, but poorly on independent test patterns drawn from the same distribution as the training patterns) (see Fig. 1(c)).
Computational complexity refers to the time requirement for a learning algorithm to estimate a solution from the training patterns. Many existing learning algorithms have high computational complexity. Designing ecient algorithms for neural network learning is a very active research topic.
There are four basic types of learning rules: (i) error-correction, (ii) Boltzmann, (iii) Hebbian, and (iv) competitive learning.
3.3.1 Error-Correction RulesIn the supervised learning paradigm, the network is given a desired output for each input pattern. During the learning process, the actual output, y, generated by the network may not equal the desired output, d. The basic principle of error-correction learning rules is to use the error signal (d y) to modify the connection weights such that this error will be gradually reduced.
The well-known perceptron learning rule is based on this error-correction principle. A perceptron consists of a single neuron with adjustable weights, wj; j = 1;2; ;n, and threshold , as shown in Fig. 3. Given an input vector x = (x ;x ; ;xn)t, the net input to the neuron (before applying the threshold function) is The output y of the perceptron is +1 if v > 0, and 0 otherwise. In a two-class classication problem, the perceptron assigns an input pattern to one class if y = 1, and to the other class if y = 0. The linear equation denes the decision boundary (a hyperplane in the n-dimensional input space) which divides the space into two halves.
Rosenblatt [18] developed a learning procedure to determine the weights and threshold in a perceptron, given a set of training patterns. The perceptron learning procedure can be described as follows.
1. Initialize the weights and threshold to small random numbers.
2. Present a pattern vector (x ;x ; ;xn)t, and evaluate the output of the neuron.
3. Update the weights according to wj(t + 1) = wj(t) + (d y)xj; where d is the desired output, t is iteration number, and  (0:0 <  < 1:0) is the gain (step size).
Note that learning occurs only when an error is made by the perceptron. Rosenblatt proved that if the training patterns are drawn from two linearly-separable classes, then the perceptron learning procedure will converge after a nite number of iterations. This is the well known perceptron convergence theorem. In practice, one does not know whether the patterns are linearly separable or not. Many variations of this learning algorithm have been proposed in the literature [8]. Other activation functions can also be used, which lead to dierent learning characteristics. However, a single layer perceptron can only separatelinearly separable patterns, as long as a monotonic activation function is used.
The well-known backpropagation learning algorithm (described in section 4) is also based on the error-correction principle.
3.3.2 Boltzmann LearningBoltzmann machines are symmetric recurrent networks consisting of binary units (+1 for on" and -1 for o"). By symmetric, we mean that the weight on the connection from unit i to unit j is equal to the weight on the connection from unit j to unit i (wij = wji). Only a portion of neurons, visible neurons, interact with the environment, the rest, called hidden neurons, do not interact. Each neuron is a stochastic unit which generates an output (or state) according to the Boltzmann distribution of statistical mechanics. Boltzmann machines operate in two modes: (i) Clamped mode in which visible neurons are clamped onto specic states determined by the environment; and (ii) Free-running mode in which both the visible and hidden neurons are allowed to operate freely.
Boltzmann learning is a stochastic learning rule derived from information-theoretic and thermodynamic principles (see [2]). The objective of Boltzmann learning is to adjust the connection weights such that the states of visible units satisfy a particular desired probability distribution. According to the Boltzmann learning rule, the change in the connection weight
wij = (ij ij); where  is the learning rate, and ij and ij are the correlations between the states of unit i and unit j when the network operates in the clamped mode and free-running mode, respectively. The values of ij and ij are usually estimated from Monte Carlo experiments which are extremely slow.
Boltzmann learning can be viewed as a special case of error-correction learning in which error is measured not as the direct dierence between the desired output and actual output, but as the dierence between the correlations between the outputs of two neurons under two operating conditions (clamped and free-running).
3.3.3 Hebbian RuleThe oldest learning rule is Hebb's postulate of learning [7]. It was proposed by Hebb based on the following observation from neurobiological experiments: if neurons on both sides of a synapse are activated synchronously and repeatedly, then the strength of that synapse is selectively increased [6].
Mathematically, the Hebbian rule can be described as wij(t + 1) = wij(t) + yj(t)xi(t); where xi and yj are the output values of neurons i and j, respectively, which are connected by the synapse wij, and  is the learning rate. Note that xi is the input to the synapse.
An important property of this rule is that learning is done locally, i.e., the change of the synapse weight depends only on the activities of the two neurons connected by it. This signicantly simplies the complexity of the learning circuit in a VLSI implementation.
A single neuron trained using the Hebbian rule exhibits an orientation selectivity. Fig. 6 demonstrates this property. The points depicted in Fig. 6 are drawn from a 2-dimensional Gaussian distribution and used for training a neuron. The weight vector of the neuron is initialized to w as shown in the gure. As the learning proceeds, the weight vector moves closer and closer to the direction w of maximal variance in the data. In fact, w is the eigenvector of the covariance matrix of the data corresponding to the largest eigenvalue.
Figure 6: Orientation selectivity of a single neuron trained using the Hebbian rule.
3.3.4 Competitive Learning RulesUnlike Hebbian learning (where multiple output units can be red simultaneously), compet- itive learning has all the output units compete among themselves for activation. As a result of such a competition, only one output unit, is active at any given time. This phenomenon is often known as winner-take-all. Competitive learning has been found to exist in biological neural networks [6].
The outcome of competitive learning is often a clustering or categorization of the input data. Similar patterns are grouped by the network and represented by a single unit. This grouping process is done by the network automatically based on the correlations in the data.
The simplest competitive learning network consists of a single layer of output units as shown in Fig. 5. Each output unit i in the network connects to all the input units (xi's) via weights, wij, j = 1;2; ;d. Each output unit also connects to all the other output units via inhibitory weights, but has a self-feedback with an excitatory weight. As a result of competition, only the unit i with the largest (or the smallest) net input becomes the winner, i.e., wi x wi x; i; or wi x wi x ; i: When all the weight vectors are normalized, these two inequalities are equivalent.
A simple competitive learning rule can be stated as follows.
Note that only the weights of the winner unit get updated. The eect of this learning rule is to move the stored pattern in the winner unit (weights) a little bit closer to the input pattern. A geometric interpretation of competitive learning is demonstrated in Fig. 7.
In this example, we assume that all the input vectors have been normalized to have unit length. They are depicted as black dots in Fig. 7(a). The weight vectors of the three units are randomly initialized. Their initial positions and nal positions on the sphere after competitive learning are shown as crosses in Figs. 7(a) and 7(b), respectively. As we can see from Fig. 7, each of the three natural groups (clusters) of patterns has been discovered by an output unit whose weight vector points to the center of gravity of the discovered group.
Figure 7: An example of competitive learning: (a) before learning; (b) after learning.
One can see from the competitive learning rule that the network will never stop learning (updating weights) unless the learning rate  is zero. It is possible that a particular input pattern may re dierent output units at dierent iterations during learning. This brings up the stability issue of a learning system. A learning system is said to be stable if no pattern in the training data changes its category after a nite number of learning iterations. One way of achieving stability is to force the learning rate to decrease gradually as the learning process proceeds, and so it eventually approaches zero. However, this articial freezing of learning causes another problem termed plasticity, which is dened as the ability to adapt to new data. This is known as Grossberg's stability-plasticity dilemma in competitive learning.
The most well-known example of competitive learning is vector quantization for data compression. Vector quantization has been widely used in speech and image processing for ecient storage, transmission and modeling. The goal of vector quantization is to represent a set or distribution of input vectors by a relatively small number of prototype vectors (weight vectors), or a codebook. Once a codebook has been constructed and agreed upon, we need only transmit or store the index of the corresponding prototype to the input vector. Given an input vector, its corresponding prototype can be found through searching for the nearest prototype in the codebook.
3.3.5 Summary of Learning AlgorithmsVarious learning algorithms and their associated network architectures are summarized in Table 2. However, this is by no means an exhaustive list of the learning algorithms available in the literature. Both supervised and unsupervised learning paradigms employ learning rules based on error-correction, Hebbian, and competitive learning. Learning rules based on error-correction can be used for training feedforward networks, while Hebbian learning rules have been used for all types of network architectures. However, each learning algorithm is designed for training a specic network architecture. Therefore, when we talk about a learning algorithm, it is implied that there is a particular network architecture associated with it. Each learning algorithm is able to perform well on at most a few tasks. The last column of table 2 lists a number of tasks that each learning algorithm can perform. Due to space limitations, we will not discuss some of the other algorithms, including ADALINE, MADALINE [13], linear discriminant analysis [10], ART2, ARTMAP [4], Sammon's projec- tion [10], and principal component analysis [8]. Interested readers can further consult the corresponding references (in order to reduce the size of the bibliography, this article does not always cite the rst paper that proposed a particular algorithm).
4 Multilayer Feedforward NetworksFig. 8 shows a typical 3-layer perceptron. In general, a standard L-layer feedforward network1 consists of one input stage, (L 1) hidden layers, and one output layer of units which are successively connected (fully or locally) in a feedforward fashion with no connections between units in the same layer and no feedback connections between layers.
The most popular class of multi-layer feedforward networks is multi-layer perceptrons in Learning Algorithm Perceptron learning algorithms pattern classication function approximation ADALINE & MADALINE prediction, control Boltzmann learning algorithm pattern classication Linear discriminant analysis pattern classication Learning vector quantization pattern classication Sammon's projection Principal component analysis Associative memory learning associative memory Vector quantization pattern classication RBF learning algorithm function approximation prediction, control Table 2: Well-known learning algorithms.
which each computational unit employs either the thresholding function or the sigmoid func- tion. Multi-layer perceptrons are capable of forming arbitrarily complex decision boundaries and can represent any Boolean function [16]. The development of the back-propagation learn- ing algorithm for determining weights in a multi-layer perceptron has made these networks the most popular among researchers as well as users of neural networks.
We denote wij l as the weight on connection between the ith unit in layer (l 1) to jth unit in layer l.
Let (x ;d );(x ;d ); ;(x p ;d p ) be a set of p training patterns (input-output pairs), where x i Rn is the input vector in the n-dimensional pattern space, and d i [0;1]m, a m-dimensional hyper-cube. For classication purposes, m is the number of classes.
The squared-error cost function, which is most frequently used in the ANN literature, is The back-propagation algorithm [19] is a gradient-descent method to minimizethe squared- Figure 8: A typical 3-layer feedforward network architecture.
error cost function in Equation (2), and is given below.
1. Initialize the weights to small random values; 2. Randomly choose an input pattern x  ; 3. Propagate the signal forward through the network; 4. Compute Li in the output layer (oi = yLi) where hli represents the net input to the ith unit in the lth layer, and g is the derivative of the activation function g.
5. Compute the deltas for the preceding layers by propagating the errors back- for l = (L 1); ;1.
6. Update weights using 7. Go to step 2 and repeat for the next pattern until the error in the output layer is below a pre-specied threshold or a maximum number of iterations is reached.
A geometric interpretation (adopted and modied from [13]) shown in Fig. 9 can help explicate the role of hidden units (with the threshold activation function). Each unit in the rst hidden layer forms a hyper-plane in the pattern space; boundaries between pattern classes can be approximated by hyper-planes. A unit in the second hidden layer forms a hyper-region from the outputs of the rst-layer units; a decision region is obtained by per- forming an AND" operation on hyperplanes. The output-layer units combine the decision regions made by the units in the second hidden layer by performing logical OR" operations.
Remember that this scenario is depicted only to help us understand the role of hidden units.
Their actual behavior, after we train the network, could be dierent from this. A two-layer network can form more complex decision boundaries than what is depicted in Fig. 9. More- over, multilayer perceptrons with sigmoid activation functions can form smooth decision boundaries rather than piece-wise linear boundaries.
Figure 9: A geometric interpretation of the role of hidden units.
A special class of multi-layer feedforward networks is the Radial Basis Function (RBF) network [6], a two-layer network. Each unit in the hidden layer employs a radial basis function, such as a Gaussian kernel, as the activation function. The radial basis function (or kernel function) is centered at the point specied by the weight vector, associated with the unit. Both the positions and the widths of these kernels must be learned from the training patterns. The number of kernels in the RBF network is usually much less than the number of training patterns. Each output unit implements a linear combination of these radial basis functions. From the point of view of function approximation, the hidden units provide a set of functions that constitute an arbitrary basis" for representing input patterns in the space spanned by the hidden units.
There are a variety of learning algorithms for the RBF network [6]. The basic algorithm employs a two-step learning strategy (hybrid learning): estimation of kernel positions and kernel widths using some unsupervised clustering algorithm, followed by a supervised least mean square (LMS) type of algorithm to determine the connection weights to the output layer. Since the output units are linear, a non-iterative algorithm can be used. After this initial solution is obtained, a supervised gradient-based algorithm can be used to rene the network parameters.
This hybrid learning algorithm for training the RBF network converges much faster than the backpropagation algorithm for training multi-layer perceptrons. However, for many problems, the RBF network often involves a larger number of hidden units compared with a multi-layer perceptron. This implies that the run-time (after training) speed of the RBF network is often slower than the run-time speed of a multi-layer perceptron. The eciencies (error versus network size) of the RBF network and the multi-layer perceptron are, however, problem-dependent. It has been shown that the RBF network has the same asymptotic approximation power as a multi-layer perceptron.
There are many issues in designing feedforward networks. These issues include: (i) how many layers are needed for a given task?; (ii) how many units per layer?; (iii) what can we expect a network to generalize on data not included in the training set?; and (iv) how large should the training set be for good" generalization? Although multilayer feedforward networks using backpropagation have been widely used for classication and function ap- proximation (see [8]), many design parameters still have to be determined by trial-and-error.
Existing theoretical results provide only very loose guidelines for selecting these parameters in practice.
5 Kohonen's Self-Organizing MapsKohonen's self-organizing map (SOM) [11] has the desirable property of topology preser- vation which captures an important aspect of the feature maps in the cortex of the more developed animal brains. By a topology preserving mapping, we mean that nearby input patterns should activate nearby output units on the map. The basic network architecture of Kohonen's SOM is shown in Fig. 5. It basically consists of a two-dimensional array of units, each of which is connected to all d input nodes. Let wij denote the d-dimensional vector associated with the unit at location (i;j) of the 2-D array. Each neuron computes the Euclidean distance between the input vector x and the stored weight vector wij.
Kohonen's SOM is a special type of competitive learning network which denes a spatial neighborhood for each output unit. The shape of the local neighborhood can be either square, rectangular, or circular. Initial neighborhood size is often set to 1/2 to 2/3 of the network size, and shrinks with time according to some schedule (e.g., an exponentially decreasing function). During competitive learning, all the weight vectors associated with the winner and its neighboring units are updated.
Kohonen's SOM learning algorithm can be described as follows.
1. Initialize weights to small random numbers; set initial learning rate and neigh- 2. Present a pattern x, and evaluate the network outputs; 3. Select the unit (ci;cj) with the minimum output: 4. Update all the weights according to the following learning rule; ij(t) + (t)[x(t) ij(t)]; if (i; j) where Nc c (t) is the neighborhood of the unit (ci;cj) at time t, and (t) is the learning rate.
5. Decrease the value of (t) and shrink the neighborhood Nc c (t); 6. Repeat steps 2 { 5 until the change in weight values is less than a pre-specied threshold, or a maximum number of iterations is reached.
Kohonen's SOM can be used for projection of multivariate data, density approximation, and clustering. Some successful applications of Kohonen's SOM can be found in the areas of speech recognition, image processing, robotics, and process control [8]. The design param- eters include the dimensionality of the neuron array, number of neurons in each dimension, shape of neighborhood, shrinking schedule of the neighborhood, and the learning rate.
6 Adaptive Resonance Theory ModelsRecall that an important issue in competitive learning is the stability-plasticity dilemma.
How do we learn new things (plasticity) and yet retain the stability which ensures that the existing knowledge is not erased or corrupted? Carpenter and Grossberg's Adaptive Resonance Theory models (ART1, ART2, and ARTMAP) were developed in an attempt to overcome this dilemma [4]. The basic idea of these models is as follows. The network has a sucient supply of output units, but they are not used until deemed necessary. A unit is said to be committed (uncommitted) if it is (is not) being used. The learning algorithm updates the stored prototypes of a category only if the input vector is suciently similar to them. An input vector and a stored prototype are said to resonate when they are suciently similar.
The extent of similarity is controlled by a vigilance parameter, , with 0 <  < 1, which also determines the number of categories. When the input vector is not suciently similar to any existing prototype in the network, a new category is created and an uncommitted unit is assigned to this new category with the input vector as the initial prototype. If no such uncommitted unit exists, then a novel input generates no response.
Competitive (output) Layer Comparison (input) Layer Figure 10: ART1 network.
We present only ART1 which takes binary (0/1) input to illustrate the model. Fig. 10 shows a simplied diagram of the ART1 architecture (see [8]). It consists of two layers of units which are fully connected. A top-down weight vector wj is associated with unit j in the input layer, and bottom-up weight vector wi is associated with output unit i; wi is the normalized version of wi.
where " is a small number which is used to break the ties in selecting the winner. The top-down weight vectors, wj's, store the prototypes of clusters. The role of normalization is to prevent prototypes with a long vector length from dominating prototypes with a short vector length. Given an N-bit input vector x, the output of the auxiliary unit A is given by A = Sgn = (Xxj N XOi 0:5); where Sgn = (x) is the signum function which produces +1 if x 0 and 0 otherwise, and the output of an input unit is given by Vj = Sgn = (xj + XwjiOi + A 1:5) if no output Oj is on"; i wjiOi; otherwise: A reset signal R is generated only when the similarity is less than the vigilance level.
The ART1 learning algorithm is described below.
1. Initialize wij = 1, for all i;j. Enable all the output units.
2. Present a new pattern x.
3. Find the winner unit i among the enabled output units 4. Vigilance test If r  (resonance), goto Step 5. Otherwise, disable unit i and goto Step 3 (until all the output units are disabled).
5. Update the winning weight vector wi , enable all the output units and goto
wji = (Vj wji ): 6. If all the output units are disabled, select one of the uncommitted output units and set its weight vector to x. If there is no uncommitted output unit (capacity is reached), the network rejects the input pattern.
The ART1 model is able to create new categories and to reject an input pattern when the network reaches its capacity. However, the number of categories discovered in the input data by ART1 is sensitive to the vigilance parameter.
7 Hopeld NetworkThe Hopeld network uses a network energy function as a tool for designing recurrent net- works and for understanding its dynamic behavior [9]. Hopeld's formulation made explicit the principle of storing information as dynamically stable attractors, and popularized the use of recurrent networks for associative memory and for solving combinatorial optimization A Hopeld network with N units has two versions: binary and continuous valued net- works. Let vi be the state or output of the ith unit. For binary networks, vi is either +1 or -1, but for continuous networks, vi can be any value between 0 and 1. Let wij be the synapse weight on the connection from unit i to unit j. In Hopeld network, wij = wji; i;j (symmetric network), and wii = 0; i (no self-feedback connections). The network dynamics for the binary Hopeld network is vi = Sgn(Xwijvj i): The dynamic update of network states in Equation (4) can be carried out in at least two ways: synchronously and asynchronously. In a synchronous updating scheme, all the units are updated simultaneously at each time step. A central clock is therefore required to synchronize the process. On the other hand, an asynchronous updating scheme selects one unit at a time, and updates its state. The unit for updating can be chosen randomly.
The energy function of the binary Hopeld network in a state v = (v ;v ; ;vN)T is The central property of the energy function is that as the state of network evolves according to the network dynamics (Eq. (4)), the network energy always decreases, and eventually reaches a local minimum point (attractor) where the network stays with a constant energy.
Suppose a set of patterns is stored in these attractors of a network. Then it can be used as an associative memory. Any pattern present in the basin of attraction of a stored pattern can be used as an index to retrieve it.
An associative memory usually operates in two phases: storage and retrieval. In the storage phase, the weights in the network are determined in such a way that the attractors of the network memorize a set of p N-dimensional patterns x ;x ; ;xp to be stored. A generalization of the Hebbian learning rule can be used for setting connection weights wij.
In the retrieval phase, the input pattern is used as the initial state of the network, and the network evolves according to the network dynamics. A pattern is produced (or retrieved) when the network reaches an equilibrium state.
How many patterns can be stored in a network with N binary units? In other words, what is the memory capacity of a network? Note that the capacity is nite because a network with N binary units has a maximum of 2N distinct states, and not all of them are attractors.
Moreover, not all the attractors (stable states) can store useful patterns. There also existspurious attractors which store patterns dierent from any of the patterns in the training It has been shown that the maximumnumber of random patterns that a Hopeld network can store is Pmax 0:15N. If the number of stored patterns p < 0:15N, then a nearly perfect recall can be achieved. If memory patterns are orthogonal vectors instead of random patterns, then more patterns can be stored. But, the number of spurious attractors increases as p reaches the capacity limit. Several learning rules have been proposed for increasing the memory capacity of Hopeld networks (see [8]). Note that we require N connections in the network to store p N-bit patterns.
Hopeld networks always evolve in the direction that leads to a lower network energy.
This implies that if a combinatorial optimization problem can be formulated as minimizing the network energy, then the Hopeld network can be used to nd the optimal (or subopti- mal) solution by letting the network evolve freely. In fact, any quadratic objective function can be rewritten in the form of Hopeld network energy. For example, the classical traveling salesperson problem can be formulated as a network energy minimization problem.
8 ApplicationsWe have discussed a number of important ANN models and learning algorithms proposed in the literature. These ANN models and learning algorithms have been widely used for solving the seven classes of problems that are described in Section 1. In Table 2, we show the typical tasks that each of the ANN models and learning algorithms is particularly suitable for. It is important to keep in mind that in order to successfully apply an ANN model and learning algorithm to a real-world problem, one must deal with a number of design issues, including network model, network size, activation function, learning parameters, and number of trainign samples. In this section, we take one of the most successful applications of ANNs,Optical Character Recognition (OCR), as an example to illustrate how multilayer feedforward networks are used in practice.
OCR deals with the problem of processing a scanned image of text and transcribing it into a machine readable form. In this section we will outline the basic components of OCR and explain how ANNs are used for character classication.
An OCR system usually consists of the following modules: (i) preprocessing, (ii) segmen- tation, (iii) feature extraction, (iv) classication, and (v) contextual processing. A paper document is scanned to produce a gray level or binary (black-and-white) image. In the preprocessing stage, ltering is applied to remove noise, and text areas are located and con- verted to a binary image using either a global or a local adaptive thresholding method. In the segmentation step, the text image is separated into individual characters. This is a particularly dicult task with handwritten text where there is a proliferation of touching characters. One eective technique is to break the composite pattern into smaller patterns (over-segmentation) and nd the correct character segmentation points using the output of a pattern classier.
Recognizing segmented characters is not an easy task because there are many dierent writing styles, dierent degrees of slant, skew, and noise level. This is evident from Fig. 11 which shows the size-normalized character bitmaps of a sample set from the NIST hand-print character database [21].
There are two main schemes for using ANNs in an OCR system as shown in Fig. 12.
The rst scheme employs an explicit feature extractor (not necessarily a neural network).
Figure 11: A sample set of characters in the NIST data.
For instance, contour direction features are used in Fig. 12. The extracted features are passed to the input stage of a multilayer feedforward network (e.g., [17]). This scheme is very exible in incorporating a large variety of features. The other scheme does not explicity extract features from the raw data. The feature extraction implicitly takes place within the intermediate stages of the ANN. A nice property of this scheme is that feature extraction and classication are integrated and trained simultaneously to produce optimal" classication results. It is not clear whether the types of features which can be extracted by this integrated architecture are the most eective ones for character recognition. Moreover, this scheme requires a much larger network than the rst scheme.
A typicalexampleof this integrated feature extraction-classication schemeis the network developed by Le Cun et al. [12] for zip-code recognition. A 16 16 normalized gray level image is presented to a feedforward network with three hidden layers. The units in the rst hidden layer are locally connected to the units in input layer, forming a set of local feature maps. The second hidden layer is constructed in a similar way as the rst hidden layer.
Each unit in the second hidden layer also combines local information coming from feature maps in the rst hidden layer. The activation level of an output unit can be interpreted as an approximation of the a posteriori probability of belonging to a particular class given the input pattern. The output categories are ordered according to activation levels and passed to the post-processing stage. In the post-processing stage, contextual information is Contour direction features Recognizedtext in ASCII Figure 12: Two schemes for using ANNs in an OCR system.
exploited to update the output of the classier. This could, for example, involve looking up a dictionary of admissible words, or utilizing syntactic constraints present, for example, in phone numbers or social security numbers.
How good are ANNs for OCR? ANNs are found to work very well in practice. However, there is no conclusive evidence about ANN's superiority over conventional statistical pattern classiers. At the First Census Optical Character Recognition System Conference held in 1992 [21], more than 40 dierent handwritten character recognition systems were evaluated based on their performance on a common database. The top ten performers among them used either some type of multilayerfeedforward network or a nearest neighbor-based classier.
ANNs tend to be superior in terms of speed and memory requirements compared to nearest neighbor methods. Unlike the nearest neighbor methods, classication speed using ANNs is independent of the size of the training set. The recognition accuracies of the top OCR systems on the NIST isolated (pre-segmented) character data were above 98% for digits, 96% for upper-case characters, and 87% for lower-case characters. One conclusion drawn from the test is that the recognition performance of OCR systems on isolated characters is comparable to the human performance. However, humans still outperform OCR systems on unconstrained and cursive handwritten documents.
9 Concluding RemarksDevelopments in ANNs have stimulated a lot of enthusiasm and criticism as well. Many comparative studies provide an optimistic outlook for ANNs, while others oer a pessimistic view. For many tasks, such as pattern recognition, no single approach dominates the other.
The choice of the best technique should be driven by the nature of the given application.
We should try to understand the capacities, assumptions, and applicability of various ap- proaches, and maximally exploit the complementary advantages of these approaches in order to develop better intelligent systems. Such an eort may lead to a synergistic approach which combines the strengths of ANNs and other approaches in order to achieve a signicantly bet- ter performance for challenging problems. As Minsky [15] has observed, the time has come to build systems out of diverse components. In such a synergistic approach, not only are individual modules important, but we also need a good methodology for integration. It is clear that communication and cooperative work between researchers working in ANNs and other disciplines will not only avoid repetitious work but, more importantly, will stimulate and benet individual disciplines.
Acknowledgment: The authors would like to thank Richard Casey (IBM Almaden), Pat Flynn (Washington State Univ.), William Punch, Chitra Dorai and Kalle Karu (Michigan State Univ.), Ali Khotanzad (Southern Methodist Univ.), and Ishwar Sethi (Wayne State Univ.) for their many useful suggestions.
[1] DARPA Neural Network Study. AFCEA International Press, 1988.
[2] James A. Anderson and Edward Rosenfeld. Neurocomputing: Foundations of Research.
MIT Press, Cambridge, Massachusetts, 1988.
[3] S. Brunak and B. Lautrup. Neural Networks, Compters with Intuition. World Scientic, Singapore, 1990.
[4] G. A. Carpenter and S. Grossberg. Pattern Recognition by Self-Organizing Neural Net- works. MIT Press, Cambridge, MA, 1991.
[5] J. Feldman, M. A. Fanty, and N. H. Goddard. Computing with structured neural networks. IEEE Computer, pages 91{103, March 1988.
[6] S. Haykin. Neural Networks: A Comprehensive Foundation. MacMillan College Pub- lishing Company, New York, 1994.
[7] D. O. Hebb. The Organization of Behavior. Wiley, New York, 1949.
[8] J. Hertz, A. Krogh, and R. G. Palmer. Introduction to the Theory of Neural Computa- tion. Addison-Wesley, Redwood City, 1991.
[9] J. J. Hopeld. Neural networks and physical systems with emergent collective compu- tational abilities. In Proc. Natl. Acad. Sci. USA 79, pages 2554{2558, 1982.
[10] A. K. Jain and J. Mao. Neural networks and pattern recognition. In J. M. Zurada, R. J. Marks II, and C. J. Robinson, editors, Computational Intelligence: Imitating Life, pages 194{212. IEEE Press, New York, 1994.
[11] T. Kohonen. Self Organization and Associative Memory. Third edition, Springer-Verlag, [12] Y. Le Cun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Back-propagation applied to handwritten zipcode recognition. NeuralComputation, 1:541{551, 1989.
[13] R. P. Lippmann. An introduction to computing with neural nets. IEEE ASSP Magazine, 4 (2):4{22, Apr. 1987.
[14] W. S. McCulloch and W. Pitts. A logical calculus of ideas immanent in nervous activity.
Bulletin of Mathematical Biophysics, 5:115{133, 1943.
[15] M. Minsky. Logical versus analogical or symbolic versus connectionist or neat versus scruy. AI Magazine, 65(2):34{51, 1991.
[16] M. Minsky and S. Papert. Perceptrons: An Introduction to Computational Geometry.
MIT Press, Cambridge, MA, 1969.
[17] K. Mohiuddin and J. Mao. A comparative study of dierent classiers for handprinted character recognition. In E. S. Gelsema and L. N. Kanal, editors, Pattern Recognitionin Practice IV, pages 437{448. Elsevier Science, The Netherlands, 1994.
[18] R. Rosenblatt. Principles of Neurodynamics. Spartan Books, New York, 1962.
[19] D. E. Rumelhart and J. L. McClelland. Parallel Distributed Processing: Exploration in the Microstructure of Cognition. MIT Press, Cambridge, MA, 1986.
[20] P. Werbos. Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences. PhD thesis, Applied Mathematics, Harvard University, November 1974.
[21] R. A. Wilkinson and J. Geist et al. (eds). The rst census optical character recognition system conference. Technical report, NISTIR 4912, U.S. Department of Commerce, NIST, Gaitherburg, MD 20899, 1992.

Source: http://metalab.uniten.edu.my/~abdrahim/ntl/jain96_NN_Tutorial.pdf