FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

15
FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP Mohammad Javad Fadaeieslam

Transcript of FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Page 1: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

FOUNDATIONS OF DATA MINING

NEURAL NETWORKS, MLP

Mohammad Javad Fadaeieslam

Page 2: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THE XOR PROBLEM

2

No single straight line exists that separates the two classes.

Page 3: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

AND, OR

3

Page 4: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THE PERCEPTRON TO REALIZE OR

4

Page 5: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

SOLVING XOR USING TWO LINES

To separate the two classes Aand B in the XOR problem, afirst thought that comes tomind is to draw two, insteadof one, straight lines.

5

Page 6: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THE TWO-LAYER PERCEPTRON

6

Page 7: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THE TWO-LAYER PERCEPTRON

7

Input layer Hidden layer Output layer

A careful look at the two-layer perceptron reveals that the action of

the neurons of the hidden layer is actually a mapping of the input

space x onto a linearly separable one.

Page 8: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

CLASSIFICATION CAPABILITIES OF THE TWO-LAYER PERCEPTRON

8

Page 9: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

POLYHEDRAL FORMED BY THE

NEURONS OF THE FIRST HIDDEN LAYER

OF A MULTILAYER PERCEPTRON.

9

Page 10: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

CLASSIFICATION CAPABILITIES OF THE TWO-LAYER PERCEPTRON

The first layer of neurons divides the input l-dimensionalspace into polyhedral, which are formed by hyperplaneintersections. All vectors located within one of thesepolyhedral regions are mapped onto a specific vertex of theunit Hp dimensional Hypercube. The output neuronsubsequently realizes another hyperplane, which separatesthe hypercube into two parts, having some of its vertices onone side and some on the other side.

10

Page 11: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

CLASSIFICATION CAPABILITIES OF THE TWO-LAYER PERCEPTRON

11

Page 12: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THREE-LAYER PERCEPTRON

12

Page 13: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THREE-LAYER PERCEPTRON

13

A three-layer perceptron can separate any union of

polyhedral regions.

Page 14: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THE DRAWBACKS OF MULTILAYER BASIC

PERCEPTRON

To assume that in practice we know the regions where the data

are located and we can compute the respective hyperplane

equations analytically is no doubt wishful thinking.

All we know in practice is a set of training points with therespective class labels.

14

Page 15: FOUNDATIONS OF DATA MINING NEURAL NETWORKS, MLP

Foundations of Data Mining

THE BACKPROPAGATION ALGORITHM

The backpropagation algorithm was introduced to find

weights of neural network from data.

It works by computing the gradient of the loss function with respect to each weight by the chain rule.

15