BackProp Task

Date:
2002-11-12
Authors:
David Öggesjö <it1ogda@ituniv.se>
Henric Thisell <it2thhe@ituniv.se>
Sebastian Edman <it2edse@ituniv.se>
Source Code:
source.zip
source.tar.gz
source.tar.bz2
Printable Version:
backprop.pdf

Table of Contents
1 Introduction
1.1 Problem description
1.2 Organization
2 Kohonen Data Analyzing
2.1 Introduction to Kohonen's SOFM
2.2 Introduction to the Principal Component Analysis
2.3 The Kohonen Choices
2.4 Program Structure
Readfile.c
Kohonen.c
Changeable Parameters
3 Kohonen Data Analyzing Results
3.1 Initial Problems
How to Present the Result?
Huge Data Sets
3.2 Graphs
Animation of Kohonen's SOFM in Progress
Learning Rate and Sigma Function
Change in Weights
What does the Graphs Show?
3.3 Analyzed Data Sets
Echo
Heart-disease
IR-spectra
Isolet
Mushrooms
Nettalk
Proteins
Sonar
Wine
Vowels
4 Back Propagation Network
4.1 Introduction to the Back Propagation Algorithm
The Neuron
Feed Forward Networks
Back Propagation Training
4.2 Back Propagation Choices
4.3 Program Structure
Readfile.c
Backprop.c
Changeable Parameters
4.4 Chosen Data Sets
Wine
Proteins
Vowels
5 Back Propagation Network Results
5.1 Initial Problems
Extra '-1' Neuron
What is Considered to be a Correct Prediction?
How to Present the Result?
No Coding Structure
5.2 Wine, Graphs and Results
Percentages
5.3 Proteins, Graphs and Results
Varying the Number of Neurons
Difference Between Runs
Neuron Output Specific Graphs
Many Iterations
Percentages
5.4 Vowels, Graphs and Results
Number of Neurons
Learning Rate
Momentum
Parameters A and B
Optimized Network
RMS Error
Percentages
6 Conclusion
6.1 Kohonen's SOFM
Checking the Data Sets
Size of the Network
Calculations
6.2 Back Propagation
Number of Neurons
Learning Rate Impact on Huge Data Sets
Momentum
Optimization
Over Training
Calculation Time
7 Further Work
8 Glossary