ARTIFICIAL NEURAL NETWORKS

Outline

Introduction Computation in the brain Artificial Neuron Models Types of Neural Networks

Discussion

What tasks are machines good at doing that humans are not?

What tasks are humans good at doing that machines are not?

What does it mean to learn? How is learning related to intelligence?

What does it mean to be intelligent? Do you believe a machine will ever be built that exhibits intelligence?

If a computer were intelligent, how would you know? What does it mean to be conscious?

Can one be intelligent and not conscious or vice versa?

Types of Applications

Machine learning

Cognitive science

Neurobiology

Mathematics

Philosophy

Usage

Signal processing

Control

Robotics

Pattern recognition

Speech production

Speech recognition

Vision

Financial applications

Data compression

Game playing

Computation in the Brain

The Brain is an Information Processing

System

10 billion nerve cells (neurons)

about 10 000 synapses

Massive parallel information processing

Capabilities of the Brain

Its performance tends to degrade gracefully under partial damage In contrast with most programs / engineered systems

It can learn (reorganize itself) from experience. Partial recovery from damage is possible if healthy units

can learn to take over the functions previously carried out by the damaged areas

It performs massively parallel computations extremely efficiently Complex visual perception

Comparison with Computer

Artificial Neural Networks attempt to bring computers a little closer to the brain's capabilities by imitating certain aspects of information processing in the brain, in a highly simplified way.

A branch of Artificial Intelligence

A loosely modeled system of artificial neurons

based on the human brain / neurons

A network of many very simple processors

(units)

each possibly having a (small amount of) local

memory

Artificial Neural Networks

A neural network can be a processing device, an algorithm, or actual hardware

Training rule weights of connections are adjusted based

presented patterns

Artificial Neural Networks (Ctd.)

Unidirectional communication channels

(connections)

Carry numeric (as opposed to symbolic) data

Units operate

On local data

On inputs received via connections

Properties

Learns from examples / experience to improve their performance

to adapt changes in the environment

to deal with incomplete information or noisy data

Capability of generalization Great potential for parallelism

Computations of components: independent of each other  

Benefits

Structure

•A Neural network can be considered as a black box that is able to predict an output pattern when it recognizes a given input pattern.

•Once trained, the neural network is able to recognize similarities when presented with a new input pattern, resulting in a predicted output pattern.

Network Structure

Layered circuit of neurons Neighboring layers completely connected; no other

connections (feedforward network) Arbitrary number of hidden layers allowed, but usually 0 or 1

A Simple Artificial Neuron

•The basic computational element (model neuron) is often called a node or unit. It receives input from some other units, or perhaps from an external source. Each input has an associated weight w, which can be modified so as to model synaptic learning. The unit computes some function f of the weighted sum of its inputs

•Its output, in turn, can serve as input to other units.

•The weighted sum is called the net input to unit i, often written neti.

•Note that wij refers to the weight from unit j to unit i (not the other way around).

•The function f is the unit's activation function. In the simplest case, f is the identity function, and the unit's output is just its net input. This is called a linear unit.

Processing Information

x1 w1j

x2

xi

Yj

wij

w2jNeuron j wij xi

Weights

Output

Inputs

Summations Transfer function

Operation of a Single Neuron

Activation Functions

Step function

Changing the bias weight Wo, moves the threshold location.

Components and Structure Processing Elements Network Structure of the Network

Processing Information by the Network Inputs Outputs Weights Summation Function

Neural Network Fundamentals

The learning process of a Neural Network can be viewed as reshaping a sheet of metal, which represents the output (range) of the function being mapped.

The training set (domain) acts as energy required to bend the sheet of metal such that it passes through predefined points. However, the metal, by its nature, will resist such reshaping.

So the network will attempt to find a low energy configuration (i.e. a flat/non-wrinkled shape) that satisfies the constraints (training data).

Learning Process

Learning can be done in supervised or unsupervised training.

In supervised training, both the inputs and the outputs are provided.

The network processes the inputs and compares its resulting outputs against the desired outputs. Errors are then calculated, causing the system to

adjust the weights which control the network. This process occurs over and over as the weights are

continually tweaked.

Learning Process

Learning: Three Tasks

1. Compute Outputs

2. Compare Outputs with Desired Targets

3. Adjust Weights and Repeat the Process

Neural NetworkApplication Development

Preliminary steps of system development are done ANN Application Development Process

1. Collect Data

2. Separate into Training and Test Sets

3. Define a Network Structure

4. Select a Learning Algorithm

5. Set Parameters, Values, Initialize Weights

6. Transform Data to Network Inputs

7. Start Training, and Determine and Revise Weights

8. Stop and Test

9. Implementation: Use the Network with New Cases

Data Collection and Preparation

Collect data and separate into a training set and a test set

Use training cases to adjust the weights

Use test cases for network validation

Neural Network Architecture

Representative Architectures

Associative Memory Systems Associative memory - ability to recall complete situations

from partial information Systems correlate input data with stored information Hidden Layer Three, Sometimes Four or Five Layers

Recurrent Structure

Recurrent network (double layer) - each activity goes through the network more than once before the output response is produced

Uses a feedforward and feedbackward approach to adjust parameters to establish arbitrary numbers of categories

Example: Hopfield

Neural Network Preparation (Non-numerical Input Data (text, pictures): preparation may involve

simplification or decomposition) Choose the learning algorithm Determine several parameters

Learning rate (high or low) Threshold value for the form of the output Initial weight values Other parameters

Choose the network's structure (nodes and layers) Select initial conditions Transform training and test data to the required format

Training the Network

Present the training data set to the network

Adjust weights to produce the desired output for each of the inputs Several iterations of the complete training set to get a

consistent set of weights that works for all the training data

Learning Algorithms

Two Major Categories Based On Input Format Binary-valued (0s and 1s) Continuous-valued

Two Basic Learning Categories Supervised Learning

Inputs Outputs

Unsupervised Learning Inputs

No desired outputs System decides how to group the input data

Backpropagation

"Backwards propagation of errors". Supervised learning method

Implements Delta rule (gradient descent algorithm) It requires a teacher that knows, or can

calculate, the desired output for any given input. It is most useful for feed-forward networks

(networks that have no feedback, or simply, that have no connections that loop).

Activation function should be is differentiable.

Backpropagation Steps Present a training sample to the neural network. Compare the network's output to the desired output from that

sample. Calculate the error in each output neuron. For each neuron, calculate what the output should have been, and a

scaling factor, how much lower or higher the output must be adjusted to match the desired output. This is the local error.

Adjust the weights of each neuron to lower the local error. Assign "blame" for the local error to neurons at the previous level,

giving greater responsibility to neurons connected by stronger weights.

Repeat from step 3 on the neurons at the previous level, using each one's "blame" as its error.

Training Patterns

10 steps 1000 steps

Training Pattern

The shown pattern has not been learned yet.... the global error is still high

1021 steps

Global error is getting down as training continues.....Deformed and noisy patterns are also used in training.

4000 steps

Training Pattern

8000 steps 12000 steps

Recognition

Too much noise

Backpropagation

Backpropagation (back-error propagation) Most widely used learning Relatively easy to implement Requires training data for conditioning the

network before using it for processing other data

Network includes one or more hidden layers Network is considered a feedforward approach

Externally provided correct patterns are compared with the neural network output during training (supervised training)

Feedback adjusts the weights until all training patterns are correctly categorized

Error is backpropogated through network layers Some error is attributed to each layer Weights are adjusted A large network can take a very long time to train May not converge

Backpropagation

Testing Test the network after training Examine network performance: measure the network’s

classification ability Black box testing Do the inputs produce the appropriate outputs? Not necessarily 100% accurate But may be better than human decision makers Test plan should include

Routine cases Potentially problematic situations

May have to retrain

Unsupervised Learning

Only input stimuli shown to the network Network is self-organizing Number of categories into which the network

classifies the inputs can be controlled by varying certain parameters

Examples Adaptive Resonance Theory (ART) Kohonen Self-organizing Feature Maps

The Self-Organizing Map: An Alternative NN Architecture

Kohonen Self-Organizing Map (SOM)

Unsupervised learning

Weights self-adjust to input pattern

Topology

Unsupervised Neural Networks –Kohonen Learning

Also defined – Self Organizing Map Learn a categorization of input space Neurons are connected into a 1-D or 2-D lattice. Each neuron represents a point in N-dimensional

pattern space, defined by N weights During training, the neurons move around to try and

fit to the data Changing the position of one neuron in data space

influences the positions of its neighbors via the lattice connections

Self Organizing Map – Network Structure

All inputs are connected by weights to each neuron

size of neighbourhood changes as net learns

Aim is to map similar inputs (sets of values) to similar neuron positions.

Data is clustered because it is mapped to the same node or group of nodes

SOM-Algorithm

1. Initialization :Weights are set to unique random values

2. Sampling : Draw an input sample x and present in to network

3. Similarity Matching : The winning neuron i is the neuron with the weight vector that best matches the input vector

i = argmin(j){ x – wj }

SOM - Algorithm4. Updating : Adjust the weights of the winning neuron

so that they better match the input. Also adjust the weights of the neighbouring neurons.

∆wj = η . hij ( x – wj)

neighbourhood function : hij

over time neigbourhood function gets smaller

Result: The neurons provide a good approximation of the input space and correspond

Applications Clustering: explores the similarity between patterns and places

similar patterns in a cluster. data compression data mining.

Classification/Pattern recognition: assigns an input pattern (like handwritten symbol) to one of many classes. associative memory.

Function approximation: finds an estimate of the unknown function f() subject to noise. Various engineering and scientific disciplines

Prediction/Dynamical Systems: forecasts some future values of a time-sequenced data. Prediction differs from Function approximation by considering time factor. Here the system is dynamic and may produce different results for the same input data based on system state (time).

Neural Network types can be classified based on following attributes:• Applications

-Classification-Clustering-Function approximation-Prediction

• Connection Type- Static (feedforward)- Dynamic (feedback)

• Topology - Single layer- Multilayer- Recurrent- Self-organized

• Learning Methods- Supervised- Unsupervised

Types of Neural Networks

Neural Computing Paradigms

Decisions the builder must make Size of training and test data Learning algorithms Topology: number of processing elements and their

configurations Transformation (transfer) function Learning rate for each layer Diagnostic and validation tools

Results in the Network's Paradigm

Neural Network Software

Program in: Programming language Neural network package or NN programming tool Both

Tools (shells) incorporate: Training algorithms Transfer and summation functions

May still need to: Program the layout of the database Partition the data (test data, training data) Transfer the data to files suitable for input to an ANN tool

NN Development Tools

MATLAB NNTOOL Braincel (Excel Add-in) NeuralWorks Brainmaker PathFinder Trajan Neural Network Simulator NeuroShell Easy SPSS Neural Connector NeuroWare

Limitations ofNeural Networks

Do not do well at tasks that are not done well by people

Lack explanation capabilities Limitations and expense of hardware technology

restrict most applications to software simulations Training time can be excessive and tedious Usually requires large amounts of training and test

data

Neural Networks For Decision Support

Inductive means for gathering, storing, and using experiential knowledge

Forecasting ANN in decision support: Easy sensitivity analysis

and partial analysis of input factors The relationship between a combined expert system,

ANN and a DSS ANN can expand the boundaries of DSS

Neural Networks for Regression

Discrete Inputs

Classification

Pattern Recognition

Clustering

Case study: Assessment of Implication of Competitiveness on Human

Development of Countries via Data Envelopment Analysis and Cluster Analysis

OUTPUT- ORIENTED SUPER EFFICIENCY DEA

Calculation of countries’ efficiency scores considering WEF scores as

input and HDI Scores as output

CLUSTER ANALYSIS by SOM

Classification of the countries based on WEF and HDI scores

WEF Scores- Basic requirements - Efficiency enhancers- Innovation and sophistication factors

HDI Scores- Life expectancy at birth

- Combined gross enrollment ratio for primary, secondary and tertiary schools

- GDP / capita

Analyzing the evolution of countries in competitiveness and human development perspectives

Iterative clustering

[2 X 4] [2 X 3]

[5 X 1] [2 X 2]

Results

Changes over years

Modeling construction problems using ANN

Homework

Find three articles from your work domain utilizing ANNs to solve a problem, to help decision making, etc. Scholar.google.com Sciencedirect.com www.library.itu.edu.tr

Select one of them to criticize in one or two paragraphs. How are they using ANNs? What are the inputs and outputs? Do you think it is appropriate to use ANNs for that domain? What

other technique can be used? Submit the full paper of the selected article along with your

criticism. Submit abstracts for the other two.