Applications of Neural Networks

1
Applications of Neural Networks

• By Jeff Chen

2
Background

• What is a neuron?
• types of neurons
• the sigmoid function
• how neurons are connected.

3
Background

• What is a neural network?
• types of neural networks
• training methods
• uses of neural networks

4
Backpropogation Network

• Network by layers
• each layer has neurons that connect only to the
  layer directly above and directly below it
• each neurons has weights to its parents
• Training
• training occurs by finding the error between the
  output and the expected output
• this is propogated back through the layers,
  each neuron being assigned blame through its
  weights
• Great for function approximation when the
  function is not clear, is discontinuous, etc.

5
Competitive Network

• Not just one network
• series of networks, all competing for the data
• Training
• each network runs the input data
• the network with the closest output to the
  desired output wins
• the winnning network then adjusts itself to
  better fit the desired output
• this way, each network begins to specialize in
  one section or type of input
• Great for classifying data when characteristics
  are not obvious or are complicated.

6
Introduction

• To create a Neural Network
• Components of a Neuron
• Input
• Output
• Sigmoid
• Components of a Network
• Neurons
• Connections
• Cases

7
Optimizations

• Training methods
• Previously O(N)?
• Searches for child, then searches back for parent
• Edge class
• Eliminates searches
• Makes training time for that area only O(1) per
  neuron

8
Applications and Scope

• Predict the stock market
• Train on previous days
• Predict next day's prices
• Uses Backpropogation network

9
Applications and Scope

• Classify music
• Takes in notes and instruments
• Classifies to time period of music
• Uses Competitive network

10
Development

• Using Java 1.5
• File I/O
• Network efficiency
• Training on data sets

11
Results and Conclusions

• Backpropogation network
• Algorithm needs tweaking
• Finds patterns very quickly
• Competitive network
• Converges too quickly
• Larger purpose
• A blackbox java implementation of neural networks
• Flexible, can be adapted