Subject: Masters Project Defense
Room:    12-3235 
Date:    Jan. 4th, 1990 


                  Kohonen's Self-Organizing Algorithm

In this project the properties of Kohonen's  self-organizing feature maps are
investigated. The Kohonen network model consists of a single layer
of nodes, where the inputs are connected to all the nodes via modifiable
weights. The nodes are arranged topologically as a 2 dimensional array.  
A C program is used to simulate the network, it demonstrates
an application of unsupervised learning where the network adjusts 
naturally to closely approximate characteristics of a set of input 
vectors.
 
The modeled neural network consists of 324 nodes arranged in an 18x18
array. Each node is connected to 224 weights.  The weights are then
adjusted as input vectors are presented to the network. There are 26 input
vectors which are used to train the network, these vectors are chosen at
random as training proceeds. The input vectors represent the 26 letters
of the alphabet in 14x16 arrays (224 weights per vector). When the network
has trained it will be able to identify the characters correctly, even if 
they are corrupted with noise. Each vector will cause a small section of 
the network to respond maximally. The network will also show how similar
looking characters will naturally be adjacent to each other.

The algorithm employs two phases: in the first phase training occurs, in
the second phase a mapping occurs to see which part of the network responded
to which input vector. Even though in this project the vectors represent 
characters in 14x16 array, they could easily be from different applications 
such as speech processing