Neural Network Assignment, Winter 2003

Introduction

In all but one experiment, the Java Neural Network Simulator was used (JNNS). The Iris dataset was then converted into the required SNNS format with a Perl script and was split (66%/33%) into training and testing (except in the last experiment where other validation methods are examined).

The methodology in conducting the experiments was to observe and record the mean-squared-error (MSE) for the training set (and sometimes the test set) over a repeated series of experiments in order to measure the effects of adjusting one variable or method.

Experiments

Standard Back-propagation Learning rate experiement

Batch Vs Standard Backpropagation

Momentum rate

Hidden neuron number

Hidden neuron activation functions

Advanced Error adjustment methods

Validation techniques