0% found this document useful (0 votes)
105 views45 pages

Lect1 PDF

The document summarizes the contents of Simon Haykin's 1998 book "Neural Networks - A comprehensive foundation". The book covers topics related to neural networks including learning processes, single and multilayer perceptrons, radial basis function networks, support vector machines, and temporal processing using feedforward networks. It also discusses models of artificial neurons, common neural network architectures, and applications of neural networks.

Uploaded by

olex
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
105 views45 pages

Lect1 PDF

The document summarizes the contents of Simon Haykin's 1998 book "Neural Networks - A comprehensive foundation". The book covers topics related to neural networks including learning processes, single and multilayer perceptrons, radial basis function networks, support vector machines, and temporal processing using feedforward networks. It also discusses models of artificial neurons, common neural network architectures, and applications of neural networks.

Uploaded by

olex
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 45

Neural Networks -

A comprehensive foundation
Simon Haykin

Prentice-Hall, 1998

2nd edition

1
1. Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Contents of Haykin’s book . . . . . . . . . . . . . . . . . . 6
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1 Benefits of neural networks . . . . . . . . . . . . . 8
1.3 Models of a neuron . . . . . . . . . . . . . . . . . 10
1.4 Neural networks as directed graphs . . . . . . . . . 17
1.5 Feedback . . . . . . . . . . . . . . . . . . . . . . 22
1.6 Network Architectures . . . . . . . . . . . . . . . 25
1.7 Knowledge Representation . . . . . . . . . . . . . 32

2
1. Preface

• Artificial Neural Network:


- consists of simple, adaptive processing units, called often neurons
- the neurons are interconnected, forming a large network
- parallel computation, often in layers
- nonlinearities are used in computations
• Important property of neural networks: learning from input data.
- with teacher (supervised learning)
- without teacher (unsupervised learning)

• Artificial neural networks have their roots in:


- neuroscience
- mathematics and statistics
- computer science
- engineering
• Neural computing was inspired by computing in human brains

3
• Application areas of neural networks:

– modeling
– time series processing
– pattern recognition
– signal processing
– automatic control

• Computational intelligence

– Neural networks
– Fuzzy systems
– Evolutionary computing
∗ Genetic algorithms

4
Neural computing has many application areas in economics and management,
because a lot of data which can be used in training of the neural network
have been saved in databases.
Inputs
Analysis:
- age

- position
- consumer habits
- debts
- creditworthiness
- family
-
-

Principle of neural modeling. The inputs are known or they can be mea-
sured. The behavior of outputs is investigated when input varies.
All information has to be converted into vector form.

5
2. Contents of Haykin’s book

1. Introduction 9. Self-Organizing Maps


2. Learning Processes 10. Information-theoretic Models
3. Single Layer Perceptrons 11. Stochastic Machines and Their
Approximates Rooted in Statis-
4. Multilayer Perceptrons
tical Mechanics
5. Radial-Basis Function
12. Neurodynamic Programming
Networks
13. Temporal Processing Using Feed-
6. Support Vector Machines
forward Networks
7. Committee Machines
14. Neurodynamics
8. Principal Components Analysis
15. Dynamically Driven Recurrent
Networks
The boldfaced chapters will be discussed in this course

6
1. Introduction

Neural networks resemble the brain in


two respects:

1. The network acquires knowledge


from its environment using a
learning process (algorithm)

2. Synaptic weights, which


are interneuron connection
strenghts, are used to store the
learned information.
Fully connected 10-4-2 feedforward
network with 10 source (input) nodes,
4 hidden neurons, and 2 output neu-
rons.
7
1.1 Benefits of neural networks
1. Nonlinearity
- Allows modeling of nonlinear functions and processes.
- Nonlinearity is distributed through the network.
- Each neuron typically has a nonlinear output.
- Using nonlinearities has drawbacks, too: local minima, difficult ana-
lysis, no closed-form easy linear solutions.

2. Input-Output Mapping
- In supervised learning, the input-output mapping is learned from trai-
ning data.
- For example known prototypes in classification.
- Typically, some statistical criterion is used.
- The synaptic weights (free parameters) are modified to optimize the
criterion.

8
3. Adaptivity
- Weights (parameters) can be retrained with new data.
- The network can adapt to nonstationary environment.
- However, the changes must be slow enough.
4. Evidential Response
5. Contextual Information
6. Fault Tolerance
7. VLSI Implementability
8. Uniformity of Analysis and Design
9. Neurobiological Analogy
- Human brains are fast, powerful, fault tolerant, and use massively
parallel computing.
- Neurobiologists try to explain the operation of human brains using
artificial neural networks.
- Engineers use neural computation principles for solving complex
problems.
9
1.3 Models of a neuron
A neuron is the fundamental information processing unit of a neural network.

Bias
bk

X1 w k1

Activation
X2 w k2 function
Output

Σ φ (.) yk

Summing
junction
Xm w km

Input Synaptic
signals weights

The neuron model consists of three (or four) basic elements


10
Bias
1. A set of synapses or connecting bk

links: X1 w k1
- Characterized by weights
(strengths). X2 w k2
Activation
function
Output
- Let xj denote a signal at the input Σ φ (.) yk

of synapse j.
Summing
- When connected to neuron k, xj is Xm w km
junction

multiplied by the synaptic weight wkj .


- weights are usually real numbers. Input
signals
Synaptic
weights

2. An adder (linear combiner):


- Sums the weighted inputs wkj xj .

3. An activation function:
- Applied to the output of a neuron, limiting its value.
- Typically a nonlinear function.
- Called also squashing function.

4. Sometimes a neuron includes an externally applied bias term bk .


11
Mathematical equations describing neuron k:
m
X
uk = wkj xj , (1)
j=1

yk = ϕ(uk + bk ). (2)
Here:
- uk is the linear combiner output;
- ϕ(.) is the activation function;
- yk is the output signal of the neuron;
- x1 , x2 , . . . , xm are the m input signals;
- wk1 , wk2 , . . . , wkm are the respective m synaptic weights.

A mathematically equivalent representation:


- Add an extra synapse with input x0 = +1 and weight wk0 = bk .

12
- The equations are now slightly simpler:
m
X
vk = wkj xj , (3)
j=0

yk = ϕ(vk ). (4)

13
Typical activation functions

1. Threshold function φ(v) = 1, v ≥ 0; φ(v) = 0, if v < 0

2. Piecewise-linear function: Saturates at 1 and 0

14
3. Sigmoid function

• Most commonly used in neural networks


• The figure shows the logistic function defined by
φ(v) = 1+e1−av
• The slope parameter a is important
• When a → ∞, the logistic sigmoid approaches the thres-
hold function (1.)
• Continuous, balance between linearity and nonlinearity
• φ(v) = tanh(av) allows the activation function to have
negative values

15
Stochastic model of a neuron

• The activation function of the McCulloch-Pitts early neuronal model


(1943) is the threshold function.
• The neuron is permitted to reside in only two states, say x = +1 and
x = −1.
• In the stochastic model, a neuron fires (switches its state x) according
to a probability.
• The state is x = 1 with probability P (v)
The state is x = −1 with probability 1 − P (v)
• A standard choice for the probability is the sigmoid type function
1
P (v) =
1 + exp(−v/T )

• Here T is a parameter controlling the uncertainty in firing, called pseu-


dotemperature.
16
1.4 Neural networks as directed graphs
• Neural networks can be represented in terms of signal-flow graphs.

• Nonlinearities appearing in a neural network cause that two different


types of links (branches) can appear:

1. Synaptic links having a linear input-output relation: yk = wkj xj .


2. Activation links with a nonlinear input-output relation:
yk = ϕ(xj ).

17
Signal-flow graphs

• Signal-flow graph consists of directed branches

• The branches sum up in nodes

• Each node j has a signal xj

• Branch kj starts from node j and ends at node k; wkj is the synaptic
weight corresponding the strengthening or damping of signal

18
• Three basic rules:
– Rule 1.
Signal flows only to the direction of ar-
row. Signal strength will be multiplied
with strengthening factor wkj .

– Rule 2.
Node signal = Sum of incoming sig-
nals from branches

– Rule 3.
Node signal will be transmitted to
each outgoing branch; strengthening
factors are independent of node signal

19
Example: Signal flow graph of linear combination
XM
vk = wkj xj (5)
j=0 • coefficients wk0 , wk1 . . . wkM are
weights

• x0 , x1 . . . xM are input signals

• by defining
wk = [wk0 , wk1 . . . wkM ]T and
x = [x0 , x1 . . . xM ]T

vk = wk T x = xT wk (6)

• Thus rule 1 is divided into 2 parts, while the basic rules 2 and 3 for
handling signal-flow graphs remain unchanged.

• In Haykin’s book, a mathematical definition of a neural network as a


directed graph is represented on page 17.
20
• Often the signal flow inside a neuron is not considered.

• This leads to so-called architectural graph, which describes the layout


of a neural network.

• This is the typical representation showing the structure of a neural


network.

21
1.5 Feedback
• Feedback: Output of an element of a dynamic system affects to the
input of this element.

• Thus in a feedback system there are closed paths.

• Feedback appears almost everywhere in natural nervous systems.

• Important in recurrent networks (Chapter 15 in Haykin).

• Signal-flow graph of a single-loop feedback system

22
• The system is discrete-time and
linear.
• Relationships between the input signal xj (n), internal signal x0j (n),
and output signal yk (n):

yk (n) = A[x0j (n)],

x0j (n) = xj (n) + B[yk (n)]


where A and B are linear operators.

• Eliminating the internal signal x0j (n) yields

A
yk (n) = [xj (n)].
1 − AB
Here A/(1 − AB) is called the closed-loop operator of the system, and
AB the open-loop operator.
23
• Stability is a major issue in feedback systems.

• If the feedback terms are too strong, the system may diverge or become
unstable.

• An example is given in Haykin, pp. 19-20.

• Stability of linear feedback systems (IIR filters) is studied in digital


signal processing.

• Feedback systems have usually a fading, infinite memory.

• The output depends on all the previous samples, but usually the less
the older the samples are.

• Studying the stability and dynamic behavior of feedback (recurrent)


neural networks is complicated because of nonlinearities.

24
1.6 Network Architectures
• The structure of a neural network is closely related with the learning
algorithm used to train the network.

• Learning algorithms are classified in chapter 2 of Haykin.

• Different learning algorithms are discussed in subsequent chapters.

• There are three fundamentally different classes of network architectu-


res.

25
Single-Layer Feedforward Networks

• The simplest form of neural networks.

• The input layer of source nodes projects onto an output layer of


neurons (computation nodes).

• The network is strictly a feedforward or acyclic type, because there


is no feedback.

• Such a network is called a single-layer network.

26
• A single-layer network with four nodes in both the input and output
layers.

• The input layer is not counted as a layer because no computation is


performed there.

27
Multilayer Feedforward Networks

• In a multilayer network, there is one or more hidden layers.

• Their computation nodes are called hidden neurons or hidden units.

• The hidden neurons can extract higher-order statistics and acquire


more global information.

• Typically the input signals of a layer consist of the output signals of


the preceding layer only.

28
• a 9-4-2 feedforward network with 9 source (input) nodes, 4 hidden
neurons, and 2 output neurons.

Input Hidden Output


layer layer layer

• The network is fully connected: all the nodes between subsequent


layers are connected.

29
Recurrent Networks

• A recurrent neural network


has at least one feedback loop.

• In a feedforward network there


are no feedback loops.

• Recurrent network with:


- No self-feedback loops to the
“own” neuron.
- No hidden neurons.

30
• Another recurrent network which has hidden neurons.

• The feedback loops have a profound impact on the learning capability


and performance of the network.

• The unit-delay elements


result in a nonlinear dy-
namical behavior if the
network contains nonli-
near elements.

31
1.7 Knowledge Representation
• Definition: Knowledge refers to stored information or models used by
a person or machine to interpret, predict, and appropriately respond to
the outside world.

• In knowledge representation one must consider:


1. What information is actually made explicit;
2. How the information is physically encoded for subsequent use.

• A well performing neural network must represent the knowledge in an


appropriate way.

• A real design challenge, because there are highly diverse ways of repre-
senting information.

• A major task for a neural network: learn a model of the world (envi-
ronment) where it is working.

32
• Two kinds of information about the environment:

1. Prior information = the known facts.


2. Observation (measurements). Usually noisy, but give examples
(prototypes) for training the neural network.

• The examples can be:


- labeled, with a known desired response (target output) to an input
signal.
- unlabeled, consisting of different realizations of the input signal.

• A set of pairs, consisting of an input and the corresponding desired


response, form a set of training data or training sample.

33
An example: Handwritten digit recognition

• Input signal: a digital image with black and white pixels.

• Each image represents one of the 10 possible digits.

• The training sample consists of a large variety of hand-written digits


from a real-world situation.

• An appropriate architecture in this case:


- Input signals consist of image pixel values.
- 10 outputs, each corresponding to a digit class.

• Learning: The network is trained using a suitable algorithm with a


subset of examples.

• Generalization: After this, the recognition performance of the network


is tested with data not used in learning.

34
Rules for knowledge representation

• The free parameters (synaptic weights and biases) represent knowledge


of the surrounding environment.

• Four general rules for knowledge representation.

• Rule 1. Similar inputs from similar classes should produce similar repre-
sentations inside the network, and they should be classified to the same
category.

• Let xi denote the column vector

xi = [xi1 , xi2 , . . . , xim ]T

35
• Typical similarity measures:

1. Reciprocal 1/d(xi , xj ) of the Euclidean distance

d(xi , xj ) = k xi − xj k

between the vectors xi and xj .


2. The inner product xTi xj between the vectors xi and xj .
- If xi and xj are normalized to unit length, then one can easily
see that
d2 (xi , xj ) = 2 − 2xTi xj .

36
3. A statistical measure: Mahalanobis distance
d2ij = (xi − mi )T C−1 (xj − mj )
Here mi = E[xi ] is the expectation (mean) of the vector (class)
xi , mj is the mean of xj , and
C = E[(xi − mi )(xi − mi )T ] = E[(xj − mj )(xj − mj )T ]
is the common covariance matrix of the classes represented by
the vectors xi and xj .
Assumption: difference of the classes is only in their means.
• Rule 2: Items to be categorized as separate classes should be given
widely different representations in the network.
• Rule 3: If a particular feature is important, there should be a large
number of neurons involved in representing it in the network.
• Rule 4: Prior information and invariances should be built into the de-
sign of a neural network.

37
• Rule 4 leads to neural networks with a specialized (restricted) struc-
ture.

• Such networks are highly desirable for several reasons:

1. Biological visual and auditory networks are known to be very


specialized.
2. A specialized network has a smaller number of free parameters.
- Easier to train, requires less data, generalizes often better.
3. The rate of information transmission is higher.
4. Cheaper to build than a more general network because of smaller
size.

38
How to Build Prior Information into Neural Network Design

• No general technique exists: ad-hoc procedures which are known to


yield useful results are applied instead.

• Two such ad-hoc procedures:

1. Restricting the network architecture through the use of local con-


nections known as receptive fields.
2. Constraining the choice of synaptic weights through the use of
weight-sharing.

• These procedures reduce the number of free parameters to be learned.

39
Illustrating the combined use of a receptive field and weight sharing. All four
hidden neurons share the same set of weights for their synaptic connections.

40
Bayesian probability theory

• Can be used for incorporating useful prior information

• Usually the data x is assumed to be generated by some model

• A generative model approach

• Prior information on the model parameters is represented by their


prior probability density p(θ)

• Bayes’ rule is then used to compute posterior probabilities:


p(x|θ)p(θ)
p(θ|x) = (7)
p(x)
where p(x) is the unconditional density function used for normalization
and p(x|θ) is the conditional probability

• Somewhere between classical estimation theory and


neural networks
41
• No simple, adaptive processing in each computational neuron

• Differs from classical estimation theory in that distributed nonlinear


network structures are used

• Mathematical analysis is often impossible

• Local minima may be a problem

• But such nonlinear distributed systems may lead to powerful represen-


tations

• Can be used for teaching MLP (multilayer perceptron) or RBF (radial


basis function) networks

• Also in unsupervised manner

42
How to Build Invariances into Neural Network Design

• Classification systems must be invariant to certain transformations


depending on the problem.

• For example, a system recognizing objects from images must be inva-


riant to rotations and translations.

• At least three techniques exist for making classifier-type neural networks


invariant to transformations.

1. Invariance by Structure
- Synaptic connections between the neurons are created so that
transformed versions of the same input are forced to produce the
same output.
- Drawback: the number of synaptic connections tends to grow
very large.

43
2. Invariance by Training
- The network is trained using different examples of the same
object corresponding to different transformations (for example
rotations).
- Drawbacks: computational load, generalization ability for other
objects.
3. Invariant feature space
- Try to extract features of the data invariant to transformations.
- Use these instead of the original input data.
- Probably the most suitable technique to be used for neural
classifiers.
- Requires prior knowledge on the problem.

• In Haykin’s book, two examples of knowledge representation are briefly


described:
1. A radar system for air surveillance.
2. Biological sonar system of echo-locating bats.

• Optimization of the structure of a neural network is difficult.

44
• Generally, a neural network acquires knowledge about the problem th-
rough training.

• The knowledge is represented by in a distributed and compact form by


the synaptic connection weights.

• Neural networks lack an explanation capability.

• A possible solution: integrate a neural network and artificial intelligence


into a hybrid system.

45

You might also like