Chapter 10 Neural Networks

Contents:

10.1 Introduction

10.1.1 Biological neurons and their models

Biological brains are excellent at recognizing patterns in a noisy image. They often learn from examples of such patterns, just as we learned to recognize letters and numbers when we were young, we didn't need any symbolic descriptions of how the letter or number must look like. On the other hand, pattern recognition using computer programs is very slow and tiring. Typed text, with a well defined and constant shape of characters can be recognized pretty well, but recognition of handwriting is hardly used.

It may be helpful to have a look at human brains to look for information which we can use in some form or another in computer programs. In the nervous system, electrical pulses that run over the axon of a nervous cell transmiting signals between nervous cells, called neurons. The pulses are made by ions in motion, with a speed of up to 100 m/s; contrary to electrical current where the electrons are in motion. Axons typically branch out, and each branch comes very close to another type of offshoot from a cell, called a dendrite. The contact point is called a synaptic gap or just synapse. An axon pulse releases chemical substances, the neurotransmitters, which clamp onto the dendrite receptors and help ions in or out of the dendrite.

Because of that, the electrical current in the neurons fluctuates, whereby the effects of several receptors and dendrites can primarily be summed up linearly. When the charge in a cell increases beyond a certain threshold value, the axon hillock will send a pulse over the axon. This happens when Na+ and K+ ions enter and leave the cell. When the concentrations of Na+, K+ and Cl- ions in the cell have been restored by a metabolic process, the Na/K pump, the entire process can repeat itself. Each neuron works independently of all other neurons; with exception of the cell's metabolic processes, the making and breaking down of the neurotransmitters, supply of building materials and the discharge of waste materials through the blood veins. The human central nervous system is composed of the bone marrow, the brainstem, the small brains and the large brains, which is itself composed of the thalamus, the hypothalamus and the cortex, and can be distinguished into 6 layers.

A simple model of a neuron is given above. The input values in enter via synapses. The influence of the synapse is given by a value sn where in is multiplied. The total value of the neuron then becomes: n sn• in. The output of the neuron is the function f released on that value and is then passed on to the other neurons. A neural network model is made up of several neurons connected together and an indication on how the model shall learn, for example by changes in the sns, the external input.

Neural Networks (NN) are thus models with properties such as calculation, recognition, association, learning, etc.:

A definition from Kohonen ( Neural Networks, Vol 1 1988):

"Artificial Neural Networks are massively parallel interconnected networks of simple ( usually adaptive) elements and their hierarchical organizations which are intended to interact with the objects of the real world in the same way as biological nervous systems do."

For neural networks many terms and their abbreviations are used, depending on which aspect one chooses: Neurocomputing, Connectionist models, Parallel Distributed Programs (PDP), (Artificial) Neural Networks (ANN or NN) and Neuro Informatics, to name just a few.

10.1.2 Computer Science research

The NN models are studied using mathematical and computer science methods, but also by evaluation or simulation of it on serial and parallel computers. This requires immense calculation capacity just to use a relatively small mode, up to one thousand neurons.

Within informatics we are searching for methods to accelerate the simulation of NNs on parallel computer systems. Also important are the resources for the (graphical) construction and description of NNs and for the visualization of the simulation and the results thereof. A lot of computer science research concentrates on the use of NNs for pattern recognition, such as for speech, handwriting and image recognition; and to control apparatuses such as robots. Also being researched is appropriate architecture for the execution of NNs, such as parallel machines and the direct implementation of NN models in VLSI hardware.

Within ITA we used a PowerXplorer parallel computer systems for research and parallel implementation of neural networks and for research into the use of NNs for satellite photos. This PowerXplorer is built up out of nodes, each composed of a calculation and a communication microprocessor and memory, of which 16 are connected to each other.

Primarily, we now use serial SUN workstations for the use of neural networks for pattern recognition. Both serial and parallel implementation use the same format for I/O data files and the storing of network parameters. Both implementations can be used for the same practical applications.

The PowerXplorer is connected to a SUN machine. The program development, compilation and linking is completely executed on the SUN. A maximum of four users can claim some nodes to execute a program.

10.1.3  Historical Overview

1943

McCulloch and Pitts laid down the basis for the design of neural networks. They showed that the neuron-type elements are capable of general calculations. They introduced the Logic Threshold Unit (LTU) with binary input and output and a threshold function with threshold   which works on the activation (as just mentioned) of the neuron.

1949
Hebb introduced "learning" with the suggestion that synapses are the location of biological learning. Synapses that are often active have a larger chance of becoming active again. This idea led to the learning rule of Hebb:

wj(t+1) = wj(t) + xj y    with a parameter

1958
Rosenblatt developed the perceptron, in was composed of one layer with LTUs and could deal with simple pattern recognition problems. Widrow introduced the Adaptive Linear Combiner and the Adaline (Adaptive linear neuron), frequently used in signal processing. He also introduced the supervised learning rule:

wj(t+1) = wj(t) +   xj (t-y ) with t the desired value of y at x

composed of xp and tp, and applied to each offering of a pattern p.

This is known as the Widrow-Hoff, the delta of the LMS (Least Mean Square) learning rule, because it can be shown that this minimizes the error E = 1/2 p (tp-yp)2. Looking at each pattern:

E / wj = ( E / y ) ( y / wj ) = - (tp-yp) (k xkwk / wj ) = - (tp-yp) xj

The learning parameter  must be small enough here, but at a small the minimization takes very long. An adequate choice and adjustment of is an art on its own, which you must learn in practice as a user of neural networks.

1969

Minsky and Papert showed that a one-layer perceptron network cannot solve certain problems. They thought that a network with more layers, to solve such problems, would not be capable of learning properly; years later this was found to be incorrect.

The decision point, there where the output goes from 0 to 1, of a LTU is given by k xkwk = , which defines a hyper-plane that is spanned in the n-dimensional space by x. With 2 inputs this becomes a line, as shown in the figure above. The desired output of AND and OR functions can be separated by a line, however those of the XOR cannot. For that we need two LTUs and one LTU that works on the output of the first two.

1975
Fukushima designed the cognitron and later the neo-cognitron, a generalization of the multi layer perceptron with "unsupervised" learning capabilities.

1982
Hopfield, a physicist, used methods from theoretical physics, especially statistical physics, to design an algorithm for selective storage and looking up of patterns in a 1-layer network.

1984
Kohonen designed the "topology preserving maps" or "self-organizing feature maps" with Learning Vector Quantization learning rules. This has successfully been applied to the translation of continuous speech to phonemes in Finish and Japanese.

1986
Rummelhart and McClelland, in their 3-part Parallel Distributed Programming, designed multi-layer feed-forward networks with back-propagation learning rules, an extension of the delta-learning rule. This got a lot of attention from psychologists. These NNs can solve any problem that can be defined formally. However, there are problems:

Despite this, it is still the most used NN model.

1987
Carpenter and Grossberg were the primary designers of the "adaptive resonance" theory (ART), an "unsupervised" learning model that is biologically plausible.

10.2 Models

10.2.1 Components of Neural Networks

Neural networks are often decomposed into the following components:

- Processing units

These neurons receive inputs from other units, calculate an output and send that to other neurons. Often there are input and output neurons, that connect it to the world outside, and internal or "hidden" neurons. The values of the signals are often restrained to the interval [0,1] or [-1,1].

- Activation state

These are internal states of neurons that change with time:  ai(t+1) = F ( ai(t), neti(t) )
where neti  is a vector that is calculated from the input values o. This is often only a scalar weighted sum of inputs, or the Euclidic distance between the input and the weighted vector:

 neti(t) = j ( wj,i (t) * oj(t) )  or  neti(t) = || w(t) - o(t) || 2 = j ( wj,i (t) - oj(t) ) 2

Often ai is also a scalar ai, that is equal to neti. But one can also model multiplicative dendrite effects with it, or time integration:   neti(t) = j (wj,i(t) * oj(t-) ) d . Also one can build in memory to model refractory periods with it.

- Output

This is an activation function:   oi (t) = f( ai(t) ) or delayed oi (t) = f( ai(t -t) )

The sigmoid function (see fig. 12.17):
o = 1 / ( 1 + exp( - (i-)) )
or functions of such sort are often used because its derivative exists and has good properties. The threshold value is often dealt with as the weight of an "artificial" neuron, which always returns the value 1.

With a Euclidic distance for neti, often a Gaussian function of neti is used. The output function can also be stochastic, the output is then 1 with a probability according to the curves above.

- Connections

Connections between neurons can be specified, often in a weighted matrix w, wi,j indicates the strength of the connections between the neurons i and j. A positive weight describes an exiting, a negative weight an inhibiting connect. A large amount of structures is possible, where often only a part of w is occupied.

- Learning rules

These describe the change in time of the parameters, usually affecting only the weighted matrix w and the parameters of the output function such as and . But also internal neuron functions (F and f) and network structures (adding neurons) can change. Often learning is an activity on its own, first a task is learned which can later be executed repetitively.

10.2.2 Structures

A basic structure is the "crossbar" switch.

Following in complexity is the use of lateral feedback. For the feedback unit the following holds:

dy/dt  = f (x, y, M, N) , the relaxation equation
dM/dt = g (x, y, M)  , the adaptation equation
dN/dt = h (y, N)  , the adaptation for feedback.

In general one can see (parts of) networks as "feed-forward" ( the output signals influence the input signals neither directly nor indirectly ) or "feedback". The feedback can occur on a local or global level.

We also concentrate on networks (for example Hopfield) which have no explicit input or output signals. Often one looks at the change in state of the neurons, from an initial to a final state. Often neurons are connected to all other neurons, the connection matrix w is then completely occupied. In this case often the following is true:  wi,i = 0 and/or wi,j = wj,i .

10.2.3 Learning Rules

There is a difference between "global" and "local" learning methods. With the first the learning is done from the outside, outside of the normal functioning of the neuron, in a separate learning phase. With the latter the learning rules are integrated in the neuron model itself. From a biological aspect these are the only realistic learning methods. Furthermore there is also a difference between the "supervised" and "unsupervised" learning, depending on whether the known answer is used during the learning proces or not.

In general, learning is a very laborious occupation, the patterns that must be learned must be offered to the net many times, sometimes up to 1 million times. The effects of the variations in the parameters of the learning process are not well known. Configuring suitable learning parameters for a given network and a given set of learning patterns occurs principally experimentally without much theoretical basis. Just keep trying, a skill that can only be learned by practicing. The same is true for the type of network and the size of it, it can hardly be predicted which is best for the particular problem one wants to solve.

Some classes of learning methods are:

- correlation

The weights are adjusted according to the rule of Hebb, proportional to the output of both neurons, for example:

 wi,j ( t+1) = wi,j ( t ) +  * oi(t) * oj(t)

We assume that the output of the neuron goes directly to another neuron, without being further processed. Often that is true because those manipulations are also placed in the calculation of neti. Many variations exist, both supervised and unsupervised.

- competition

This method focuses of two (parts of) layers, an input and an output layer. The output neurons compete until one dominates, for example because it has the highest activation. The weights to this neuron are then adjusted, for example according to the rule of Hebb. This procedure has the effect of grouping the input data into classes of similar properties. This method is used in the Adaptive Resonance Theory models of Carpenter and Grossberg.

- error correction

Here an error measurement is calculated depending on the output of the system and the desired result, for example:

pj ( o pj - t pj) 2  ( o: output, t: target, p: learning patterns, j: neurons)

This error is minimized with respect to the wij matrix, according to given rules for the adjustments of the weights. Known are the delta rule (according to Hebb) and the "back-propagation" rule where errors are calculated reversely through several layers.

- stochastically

Weights are adjusted to minimize a statistical unit, which often looks like the entropy function from the field of thermodynamics. A global minimum is searched for by training the system at decreasing values of a variable, analogous to the physical temperature.

-generation of "hidden" layers or extra neurons.

The number of units in a network is always a compromise between the generalization of many patterns and the learning of special patterns. The number of neurons required for a certain task is statistically unpredictable. Two ideas have been developed to determine the number of neurons needed by the learning process itself:
1) A minimal network is trained, starting with simple, easily generalized patterns. Then more difficult problems are viewed, if these do not fit appropriately then extra neurons are added.
2) The patterns are placed in classes and for each class a network is taught. These networks are then combined into one network. Problems arise when combining results and removing redundant parts.

Most of these methods do not have many known results. From a biological aspect, no or a restricted amount of new neurons are formed, for the synapses it looks as if these are formed. Of course this says nothing about switching existing neurons to new tasks.

10.3 Back-propagation

These are multi-layer feed-forward networks with back-propagation learning rules. Every neuron in a layer is connected to each of the neurons on the next layer, and there are no connections between two neurons in one layer. The input layer gives the original input of every pattern to every neuron in the first hidden layer. The input values are often restricted to an interval of [0,1] or [-1,1]. For most problems this doesn't matter much, but for some [-1,1] is more suitable. The neti of the hidden and output neurons is a weighted sum:
neti(t) = j ( wj,i  * oj ) + i

Here i is often interpreted as a weight to a virtual neuron with a fixed output of 1.

The output function fi of every neuron has neti as input, and it must be possible to differentiate:

is the slope, this is often constant and usually 1. Sometimes the slope is taught to each neuron individually. Multiplying  with a number coincides with dividing by all the weights and  with the same number. So actually,  can be taken fixed. However it is stated that by adjusting   learning occurs faster.

The target value is the desired output for an output neuron for a particular learning pattern. For classification purposes there is often 1 output per class, with a high target value (0.8 to 1) for input patterns in that class and a low (0 to 0.2, respectively -1.0 to -0.8) target value for the other input patterns. The output values must then be edited by another function to determine the output class. For the approximation of a function with real values the output has a low linear activation function, while the hidden layers usually do not have a linear activation function at all.

The next statement has been proven:

Given a >0 and a L2 function f: [0,1]n -> Rm, then there is a back-propagation network with one hidden layer that approximates f with a mean square error smaller than .

Essentially one hidden layer is enough, however the number of neurons in that layer can become enormous. It can be useful to try networks with more hidden layers, the total number of required neurons could be less and then the net can work and learn faster.

Learning from a back-propagation net occurs as follows. First the weights and s are initialized, usually with random values often with a dispersal  around 0. Then the learning patterns are offered and the error function:

E = 1/2pj ( o pj - t pj)2

calculated. The E(wij, j) is minimized to all wij  and j by a gradient descent method:

wij= -  E / wij  with  as the learning rate

For neuron j in the output layer the following is true:

-E / wij  =  - (E /  netj ) (netj / wij)
                = p (tpj - opj)  (opj / netj )  ohi
                =  p pj  ohi  (from the previous layer)

for tanh(): pj =  (tpj - opj) 1/2 (1-opj2) 

for logis(): pj =  (tpj - opj) opj (1-opj) 

For neuron i in the previous hidden layer, we can also determine -E / wki, keeping in mind that every oj is dependent on every ohi :

-E / wki  = p (j pj wij )  (oi /  neti )  ohk

similar to the output layer but then with (tpj - opj) replaced by: j pj wij

The errors (tpj - opj) on the output neurons are thus divided over the hidden neurons lying before it, that is why it is called back-propagation. After the input layer has been reached, the weights and the thresholds are adjusted and the learning patterns are offered again. For the thresholds  the same formulae hold because the thresholds can be interpreted as the weight to the virtual neuron always yielding an output of 1.

Often a large amount of iterations (epochs) are needed for each set of patterns. Learning continues until E is small enough, or even better yet, until the best results are attained using a separate set of test patterns. By "over-learning" or over-specialization it can happen that the E decreases on the learning data while it increases on the test data.

On the grounds of training data, the net on the left has learned the underlying trend in the curve well, the test data lies well in the neighborhood of the curve. To the right we see the results of "over-learning", by taking the net on the left and teaching it more or by using a net with more hidden neurons. The training data lie precisely on the curve, but the test data exhibits large deviations.

Afterwards a separate pattern verification set is used to see how well the net has "generalized". The learning, testing and verification set must be representative for the problem, attaining large sets of representative patterns is a problem on its own.

Next to "epoch" or "off-line" learning the weights can be changed after each pattern, named "on-line" learning and has a smaller learning rate. This is often faster (less epochs needed) and yields better results with more complex problems and when the input patterns are noisy. This is because it is easier to jump over a local minimum. If we start in the picture on the side with a gradient descent in p we reach the local minimum Ml and not the global minimum Mg.

To pass over smooth surfaces in E quickly and jump over shallow local minima, often a momentum term is used:
wij(t) = -  E / wij (t) +   wij(t -1)

The number of neurons and layers, the values of the parameters (and possibly adjustments during the learning process) and the initialization of the network must be determined experimentally.

There are many variations on the above mentioned "standard backprop", that sometimes yield better results than on-line learning and often require less iterations. Well-known are "Quickprop" and "RPROP", both regard E(wij) around the minimum as a parabola and look at the change in sign of  E / wij to determine wij. They have 3 respectively 5 learning parameters of which the recommended ones often yield good results.

Learning patterns that have a "good" ( tj-oj < ) output, leave out some epochs, and thereby speed up the learning process. The ßj can also be learned for each neuron. For the output neurons the following is true for tanh():
 -E / j = p (tpj - opj)  1/2 (1-opj2) netj

10.4 Kohonen networks

With Kohonen networks, every neuron has a weighted vector w with the same dimension as the input patterns. For each supplied pattern p every neuron determines: || xp -w || followed by determining the winning neuron m with the lowest match value. Weights are then adjusted depending on:

With the Self Organizing Maps or Topology Preserving Maps the neurons are ordered in a 2-D matrix. Next to the winning neuron, neurons with a distance of r(t) are also adjusted according the matrix:

wi = (t) (xp -wm) where  (t) and r(t) decrease linearly with the number of epochs.

Clusters of patterns in the feature space are placed on a cluster of one or more neurons in the matrix. This is an unsupervised learning method; one must yet determine what the belonging clusters in the matrix mean in the coinciding feature space.

The Ritter∓Schulten variation is used a lo. Here every neuron is adjusted by:

wi = (t) (xp -wm) exp (-afst(i,m)2 /  (t)2 ) with afst() the Euclidic distance in the 2-D matrix
(t) and (t) are multiplied by c < 1 per epoch

To the left we see 6 Kohonen neurons ordered in a 2 by 3, 2-D matrix. To the right the initial positions of the neurons in the weighted space are shown, which are for example chosen randomly.

To the left the positions of 6 input patterns are shown, or rather the center of the 6 classes of the input patterns. A well-trained SOM then attains the positions in the feature space such as shown on the right. These are centered on the positions of the clusters, and clusters lying close to each other in the feature space end up on neurons which also lie close to each other in the matrix. The topology of the feature space can be found in the matrix space.

Here we took randomly chosen positions in the 2-D unit-square as input patterns.

The neurons are initialized with small values around the 0. We see that the Kohonen neurons slowly develop to regular positions in the unit-square.

With Learning Vector Quantization (LVQ), several neurons from each class are taken and initialized (a supervised method), for example spread out around the average of the learning pattern. The winning neuron is adjusted:
    wi =  ±(t) (xp -wm)  (+ same class, - other class)

Besides only the winning neuron, using LVQ2 and 3, also the second best is taken into account.

10.5 Classical pattern-recognition methods

These try to find discriminating functions to divide classes:

di(x) > dj(x) for j  i when pattern x belongs to class i

With the minimum distance or minimum mean classifier, every class attains a prototype vector: mi = xchi  x / Ni

An input vector x attains class i if ||x-mi|| is minimal, the coinciding discriminating function is then: di(x) =  xTmi - 1/2 miTmi  (fig. 12.6)

The probability that pattern x comes from the class i, is indicated by: p(i | x). If a classifier decides that x comes from j while it actually comes from  i, there is a loss or error in the classification which can be indicated by Lij . Because x can belong to each of the present M classes, the average loss when x is credited to j is:

  crj(x) = k Lkj p(k | x)

Using Bayes theorem: p(| b) =  p() p(b | )  / p(b) :

   crj(x) = (1 / p(x)) k  Lkj p(x | k) p(k)

Because p(x) is a positive value this term is left out: rj(x) =k  Lkj p(x | k) p(k) (average risk or loss)

A Bayes classifier adds x to class j if rj(x) < ri(x) for all i j  and thus minimizes the total average loss caused by incorrect classifications.

In many problems the loss is 0 for a correct decision and 1 for an incorrect decision:

Lkj = 1 - kj thus: rj(x) =k  (1 -  kj ) p(x | k) p(k) = p(x) -  p(x | j) p(j)

This coincides with the decision function:   dj(x) = p(x |  j)p(j)

To be able to construct a Bayes classifier we need knowledge about the probabilities of the distributions patterns in the classes which are present, and a loss function relevant to that particular problem. Often Gaussian distributions are assumed [12.10] or other functions which can be calculated easily, and where the parameters of the functions can be determined with a certain precision from the learning patterns present.

For many medical and also other applications Lij Lji. While testing on the presence of a certain disease it is often less dramatic for a healthy person to be diagnosed as sick rather than a sick person being diagnosed as healthy. With a "false positive" decision the following and other tests will indicate the illness not to be present, with a "false negative" decision the patient goes home and the illness can develop further.

There are many, however not always known, resemblances between neural networks and classical statistical pattern recognition methods. Attractive for neural networks is that these work with examples and without the use of "higher order" knowledge of the underlying problem. When the results of the classification are sufficient (hopefully, because there is no guarantee that a problem is taught), we have enough information unless one wants to include "higher order" information (distributions, etc.) about the underlying problem. This information has been hidden in the large amount of weights and is difficult to retrieve. Using domain knowledge can often lead to better solutions.

Updated on April 4th 2003 by Theo Schouten.