You are on page 1of 51

StudentRockStars.

com

StudentRockStars.com

Introduction to Neural Networks


By Suneel

StudentRockStars.com

What is a Neural Network?


The question 'What is a neural network? is ill-posed. StudentRockStars.com -- Pinkus (1999) A method of computing, based on the interaction of multiple connected processing elements

StudentRockStars.com

Characteristics of Biological Neural Networks


1) Massive connectivity 2) Nonlinear, Parallel, Robust and Fault Tolerant 3) Capability to adapt to surroundings

StudentRockStars.com

4) Ability to learn and generalize from known examples 5) Collective behavior is different from individual behavior
Artificial Neural Networks mimics some of the properties of the biological neural networks

StudentRockStars.com

Massively parallel computers are

approaching power of human brain Can we develop neural networks which StudentRockStars.com work well on these machines? It is well known that artificial neural networks do not scale well

StudentRockStars.com

What can a Neural Net do?


Compute a known function Approximate an unknown function Pattern StudentRockStars.com Recognition Signal Processing Learn to do any of the above

StudentRockStars.com

Neural Computing
Neuron is fundamental information processing unit of brain Neuroscience
The objective is to understand the human brain Biologically realistic models of neurons Biologically realistic connection topologies

StudentRockStars.com

Neural networks
The objective is to develop computation methods Highly simplified artificial neurons Connection topologies that are aimed at computational effectiveness

StudentRockStars.com

Human Brain vs Supercomputers


Human Brain ~100 billion neurons (1011) (for comparison, a rat

cortex has about 30million) StudentRockStars.com ~1000 synapses per neuron (1014 total ) This is roughly 100 terabytes (1014) of data storage It is also capable of roughly 1000 teraflops (1015 operations per second)

StudentRockStars.com

IBM BlueGene/L Computer (DOE) has: 65,536 processors ( 1013 transistors? ) 33 terabytes RAM (1013) 137 teraflops (1014 operations per second) StudentRockStars.com NASAs Columbia computer has: 10,240 Itanium processors (1.5 GHz) 10 terabytes RAM (1013) 40 teraflops (1013 operations per second) Theoretically more capable than the brain of a monkey ... and near human capability ...

StudentRockStars.com

StudentRockStars.com

Man versus Machine (hardware)

StudentRockStars.com

Von Neumann Numbers Human brain computer (a.d. 2005) # elements 1010 - 1012 107 - 108 neurons transistors StudentRockStars.com 10 # connections / 104 element switching frequency 103 Hz 109 - 1010 Hz energy / operation power consumption reliability of system 10-16 Joule 10 Watt high 10-6 Joule 100 - 500 Watt reasonable reasonable

reliability of elements low

Man versus Machine (information processing)


Features Human Brain

StudentRockStars.com

Data analog StudentRockStars.com representation Memory distributed localized localization Control distributed localized Processing Skill acquisition parallel learning

Von Neumann computer digital

sequential programming

StudentRockStars.com

Basic Concepts
A Neural Network generally

maps a set of inputs to a set of outputs


Number of inputs/outputs is

Input 0

Input 1

...

Input n

variable

StudentRockStars.com
Output 0 Output 1

Neural Network

The Network itself is composed

of an arbitrary number of nodes with an arbitrary topology

...

Output m

StudentRockStars.com

Applications of ANNs
ANNs have been widely used in various

domains for:
Pattern recognition StudentRockStars.com Associative memory Function approximation

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

Organization of the Brain


Central nervous system Interregional circuits Local circuits Neurons StudentRockStars.com Dendritic trees Neural microcircuits Synapses Molecules

StudentRockStars.com

Biological Neuron
A neuron has four main regions: Cell body (soma) DendritesStudentRockStars.com Axon Presynaptic terminal excitatory inhibitory

StudentRockStars.com

The Biological Neuron

StudentRockStars.com

The human brain is made of about 100 billions of such neurons.

Neurons: Dendrites and Axons StudentRockStars.com (Information travels thru electrical signals)

StudentRockStars.com

Synapses

StudentRockStars.com

You are your synapses (LeDoux) Your memories, emotions, etc. are stored in

your synapses StudentRockStars.com Learning occurs via changes to the synapses Some of the synapses are set at birth, while others are trained The human ability to hear, remember, fear, or desire are all related to the synapses.

Moores Law

StudentRockStars.com

For about 50 years we have seen computer

performance double every 18 months Intel expects this to continue until at least 2011 StudentRockStars.com Number of transistors per chip area has doubled every 18 months: - 1974, Intel 8088, 29,000 transistors -2000, Intel Pentium 4, 42,000,000 transistors -2011, Intel, 20 billion transistors/chip expected (maybe 128processors per chip)

StudentRockStars.com

StudentRockStars.com

Future Computers

StudentRockStars.com

Intel expects the first petaflop (1015 ops/second)

to appear by about 2009 this is human-brainlevel computing power (it might cost $150M and require 8 MWatts of power) StudentRockStars.com By 2011 there could be several petaflop computers By 2011 there could be supercomputers several times more powerful than human brains

Artificial Neural Networks

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

Artificial Neural Networks


Layered net with n input nodes m output nodes zero or more
x1

1
x2

StudentRockStars.com

hidden layers (one shown)


m
xn

StudentRockStars.com

Advantages of ANNs
Efficient

Inherent massively parallel


Robust

Can deal with incomplete and/or noisy data


Fault-tolerant

StudentRockStars.com

Still works when part of the net fails


User-friendly

Learning instead of programming

StudentRockStars.com

Disadvantages of ANNs
Difficult to design
The are no clear design rules for arbitrary applications

Hard orStudentRockStars.com impossible to train Difficult to assess internal operation


It is difficult to find out whether, and if so what tasks are performed by different parts of the net

Unpredictable
It is difficult to estimate future network

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

Model of a Neuron
Weigthed connecting links Adder

StudentRockStars.com
Activation function

vk = wkj xj
j=1

yk = f (vk + bk)

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

Activation Functions
1) Threshold Function f(v) = 1 if v 0 = 0 otherwise 2) Piecewise-Linear Function f(v) =StudentRockStars.com 1 if v = v if > v > - = 0 otherwise 3) Sigmoid Function f(v) = 1/{1 + exp(- av)} etc..

Discrete asymmetric transfer f(c, x) = ( x > c ? 1 : 0 ) Heaviside step-function:


1.2 1 0.8 c=2

StudentRockStars.com

StudentRockStars.com 0.6
0.4 0.2 0 -0.2 -10 -5 0 5 10

StudentRockStars.com

Piecewise-Linear Transfer
f(c, z) = ( z < c ? 1 : ( z > c ? 1 : z / c ) )
1

StudentRockStars.com
0.5 0 -0.5

-1 -10 -5 0 5 10

Continuous asymmetric transfer


sigmoid function logistic function
1.2 1 0.8 c = 0.5 c = 1.0 c = 2.0

StudentRockStars.com

StudentRockStars.com 0.6
0.4 0.2 0 -0.2 -10 -5 0 5 10

StudentRockStars.com

NEURAL NETWORK VIEWED AS DIRECTED GRAPHS


A

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com
C

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

FEED BACK

StudentRockStars.com

StudentRockStars.com

Multilayer feedforward network

StudentRockStars.com
Feed Back Network

StudentRockStars.com

NETWORK ARCHITECTURES
Network architectures typically used

Single layer feedforward Multilayer feedforward StudentRockStars.com Recurrent networks, etc.

StudentRockStars.com

Different Network Topologies


Single layer feed-forward networks Input layer projecting into the output layer

StudentRockStars.com

Single layer network

Input layer

Output layer

StudentRockStars.com

Different Network Topologies


Multi-layer feed-forward networks One or more hidden layers. Input projects only from previous layers onto a layer.

StudentRockStars.com or 2-layer

1-hidden layer fully connected network Input layer Hidden Output layer layer

StudentRockStars.com

Different Network Topologies


Recurrent networks A network with feedback, where some of its inputs are connected to some of its outputs StudentRockStars.com (discrete time).
Recurrent network

Input layer

Output layer

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com Multilayered Perceptron (Contd.)

Some important issues: How big should be my network ? No specific answer is known till date. The size of the network depends on the complexity of the problem at hand and the training accuracy which is desired. A good training accuracy does not always means a good network. If the number of free parameters of StudentRockStars.com the network is almost the same as the number of data points, the network tends to memorize the data and gives bad generalization. How many hidden layers to use ? It has been proved that a single hidden layer is sufficient to do any mapping task. But still experience shows that multiple hidden layers may be sometimes simplify learning.

Artificial Intelligence(AI) and Neural Network


The AI is to development of Algorithms that require

StudentRockStars.com

machines to perform cognitive tasks.


AI systemStudentRockStars.com must be capable of 3 things

1.Store knowledge. 2.Apply the knowledge store to slove problems. 3.Acquire new knowledge through experience.

StudentRockStars.com

Representation. Reasoning. Learning

StudentRockStars.com

StudentRockStars.com

Back-Propagated Delta Rule Networks (BP)


Inputs are put

through a Hidden Layer before the StudentRockStars.com Hidden Layer ... output layer All nodes connected ... between layers
H0 H1 Hm O0 O1 Oo Output 0 Output 1

Input 0

Input 1

...

Input n

...

Output o

StudentRockStars.com

How to Build Prior Information into Neural Network Design

StudentRockStars.com

StudentRockStars.com

Block diagram of an invariant feature-space type of system

StudentRockStars.com

StudentRockStars.com

StudentRockStars.com

Thank You

You might also like