2. Text Books:
โข Neural Networks and Learning Machines โ Simon Haykin
โข Principles of Soft Computing- S.N.Shivnandam & S.N.Deepa
โข Neural Networks using Matlab- S.N. Shivanandam, S. Sumathi ,S N Deepa
4. โA neural network is a massively parallel distributed processor made
up of simple processing units that has a natural propensity for storing
experiential knowledge and making it available for use.โ
It resembles the brain in two respects:
1. Knowledge is acquired by the network from its environment
through a learning process.
2. Interneuron connection strengths, known as synaptic weights, are
used to store the acquired knowledge.
Neural Network
11. CNS- Brain and Neuron
Neuron - Structural Unit of central nervous
system i.e. Brain and Spinal Cord.
โข 100 billion neurons, 100 trillion synapses
โข Weight -1.5 Kg to 2Kg
โข Conduction Speed โ 0.6 m/s to 120 m/s
โข Power โ 20% ,20-40 Watt,10โ16 ๐ฝ
๐๐๐๐๐๐ก๐๐๐๐
โข Ion Transport Phenomenon
โข Fault tolerant
โข Asynchronous firing
โข Response time = 10โ3
sec
โThe Brain is a highly complex, non-linear and massively parallel
Computing machine.โ
๐ต๐๐๐๐๐
12. โA Neuron is a basic unit of brain that processes and transmits information.โ
Neuron
โข Dendrite: Receive signals from other
neurons
โข Soma (Cell body): Process the incoming
signals.
โข Myelin Sheath: Covers neurons and help
speed up neuron impulses.
โข Axon : Transmits the electric potential from
soma to synaptic terminal and then finally
to other neurons, muscles or glands
โข Synaptic Terminal : Release the
neurotransmitter to transmit information to
dendrites.
23. Activation Functions
Name Geometrical Shape Mathematical Expression Property
Hyperbolic Tangent
or
Bipolar sigmoid
๐ ๐ = ๐๐๐๐๐ =
๐๐โ๐โ๐
๐๐+๐โ๐
Differentiable
๐โฒ ๐ = ๐ โ ๐๐ ๐
Bipolar Hard Limit
Signum Function
๐ ๐ = ๐
Differentiable
Rectified Linear Unit
๐ ๐ = ๐๐๐(๐, ๐) Differentiable
0
1
0 ๐
๐ ๐
-1
๐ ๐
๐
๐ ๐
๐
0
24. Single Perceptron to Multiple layer of perceptron โ Historical Perspective
McCulloch Pitts (1943) โ 1st Mathematical model of neuron
Weighted sum of input signals is compared to a threshold to determine the neuron output
Hebbian Learning Algorithm -1949
Learning of weights to classify the patterns
Organization of Behaviour โ David Hebb
Frank Rosenblatt (1957) โ More Accurate Neuron Model (Perceptron)
Perceptron Learning Algorithm to find optimum weights
Perceptron Learning Algorithm
Delta rule or Widrow- Hoff Learning Algorithm
Approximate steepest descent algorithm
Least Means Square Algorithm
Adaptive Linear Neuron Network Learning
Marvin Minsky and Seymour Peppert (1969)
Limitation of perceptron in classifying Non separable Patterns
Back Propagation (1986)
Training of Multilayer of Perceptrons
๐พ๐๐๐ = ๐พ๐๐๐ + ๐๐๐
๐พ๐๐๐ = ๐พ๐๐๐ + (๐ โ ๐)๐๐
๐พ๐๐๐ = ๐พ๐๐๐ โ แ
๐ถ๐๐ญ ๐
๐=๐พ๐๐๐
25. Geometrical Significance (Hardlimit activation function)
๐ฅ1
๐ฅ2
๐ฅ๐
โฎ
๐ค1
๐ค2
๐ค๐
เท
๐ = ๐ เท
๐=๐
๐
๐๐๐๐ + ๐
เท(๐ข๐ง๐ฉ๐ฎ๐ญ)
๐
Inputs b (Bias)
Hyperplane
Activation
Function
0
1
Output
input
Hard limit Function
From Activation function, we can infer if ฯ๐=๐
๐
๐๐๐๐ + ๐ or ๐พ๐ป
๐ฟ (inner product between weight vector and input vector) is
greater than 0 for output is 1.
๐ฐ๐ก๐๐ซ๐, ๐ฐ๐๐ข๐ ๐ก๐ญ ๐ฏ๐๐๐ญ๐จ๐ซ , ๐ =
๐ฐ๐
๐ฐ๐
โฎ
๐ฐ๐ง
and input vector, ๐ฟ =
๐๐
๐๐
โฎ
๐๐
. The ฯ๐=๐
๐
๐๐๐๐ + ๐ = ๐ is equivalent to a hyperplane
boundary.
26. Geometrical Significance (Hardlimit Activation function)
2 input (๐๐ & ๐๐) โ Boundary is line (๐๐๐๐ + ๐๐๐๐+b=0) with ๐๐, ๐๐ ๐๐ ๐๐๐๐๐๐ โฅ ๐๐๐๐๐๐ ๐๐ ๐๐๐๐ .
3 input (๐๐, ๐๐ & ๐๐) โ Boundary is Plane (๐๐๐๐ + ๐๐๐๐+๐๐๐๐+b=0) with ๐๐, ๐๐, ๐๐ ๐๐ ๐๐๐๐๐๐ โฅ ๐๐๐๐๐๐ ๐๐ ๐๐๐๐๐.
>3 input (๐๐, ๐๐,โฏ ๐๐) โ Boundary is Hyperplane (๐๐๐๐ + ๐๐๐๐+โฏ+๐๐๐๐+b=0) with ๐๐, ๐๐ โฏ ๐๐ ๐๐ ๐๐๐๐๐๐ โฅ
๐๐๐๐๐๐ ๐๐ ๐๐๐๐๐๐๐๐๐๐.
๐๐๐ข๐ ๐ก๐ญ ๐๐๐๐ญ๐จ๐ซ
(๐ฐ๐, ๐ฐ๐)
๐ณ๐๐๐ ๐ฌ๐๐๐๐๐๐๐ (๐ฐ๐๐ฑ๐ + ๐ฑ๐๐ฐ๐+b=0)
Class 1
Class 2 ๐๐๐ข๐ ๐ก๐ญ ๐๐๐๐ญ๐จ๐ซ
(๐ฐ๐, ๐ฐ๐, ๐ฐ๐)
๐ท๐๐๐๐ ๐ฌ๐๐๐๐๐๐๐
(๐ฐ๐๐ฑ๐ + ๐ฑ๐๐ฐ๐++๐ฑ๐๐ฐ๐+b=0)
Class 1
Class 2
๐ฑ๐
๐ฑ๐
๐ฑ๐
๐ฑ๐
๐ฑ๐
2 Class Single Neuron Classification
33. Hebbian Learning Rule
โข Donald Hebb (Psychologist)โ The Organization of the behaviour (1949)
โข Hebbโs Postulate โ โWhen an axon of cell A is near enough to excite a cell
B and repeatedly or persistently takes part in firing it, some growth process or
metabolic change takes place in one or both cells such that Aโs efficiency, as
one of the cells firing B, is increased.โ
Mathematically,
๐พ๐๐๐= ๐พ๐๐๐ + ๐๐๐
Where ๐ฅ๐ is the ith input and ๐ฆ is output.
Bipolar inputs or outputs (-1 or +1)
Limitation โ Can classify linearly separable patterns only
39. Perceptron Learning Rule
โข Frank Rosenblatt โ (1957)
โข Key contribution - Introduction of a learning rule for training perceptron networks to
solve pattern recognition problems
โข Perceptron could even learn when initialized with random values for its weights and
biases.
โข Limitations โ Can classify only linearly separable problems.
โข Limitations were publicized in the book โPerceptrons (1969)โ by Marvin Minsky and
Seymour Peppert.
Mathematically,
๐พ๐๐๐= ๐พ๐๐๐ + (๐ โ เท
๐) ๐๐
Where, ๐ฅ๐ ๐๐ ๐๐กโ ๐๐๐๐ข๐ก, เท
๐ฆ ๐๐ ๐๐๐ก๐ข๐๐ ๐๐ ๐๐๐๐๐๐๐ก๐๐ ๐๐ข๐ก๐๐ข๐ก ๐๐๐
๐ฆ ๐๐ ๐ก๐๐๐๐๐ก ๐๐ข๐ก๐๐ข๐ก.