The Self Organizing Map: Unsupervised Competitive Learning
- Introduction
- Competitive Learning
- Covering Feature Space: A Two Dimensional Example
- The Development of Feature Maps
- References
- Slides
Introduction
In this chapter, we will look at an unsupervised architecture called the Self Organizing Map (SOM, Kohonen 1982).Figure 1: The Self Organizing Map (SOM)
The SOM consists of an input layer that is fully connected to an output layer of map nodes (see figure 1). When an input is presented the output nodes compete to represent the pattern. The node whose vector of weights is closest to the input pattern wins the competition. The winner and the units close to it (its neighbours) are then updated by moving their weight vectors closer to the input pattern. Because units near the winner are also moved, as training progresses units that are neighbours tend to come to represent similar patterns, while nodes far from each other in the map represent dissimilar patterns. If there are clusters of input patterns, then the points within a cluster will tend to activate the same output unit, while points from different clusters will be represented by separate units. The more dissimilar the clusters the further apart they will be mapped in the output layer.
In the SOM network above (figure 1) the yellow units at the bottom are inputs representing the features of members of two rival gangs, the Jets and the Sharks (see chapter 2). In our database of gang members we have information about their age, educational level, martial status and occupation. The blue output layer of units are the map nodes. We would like to use the SOM to discover any clusters of individuals. The input set contains the patterns to which we will expose the network. Each pattern represents the details of a different gang member.
Exercise 1: Select each of the input patterns in turn and click on cycle to see which of the outputs represents it. Complete the Before Learning column of the following table. Now click on the learn button (to execute 80 epochs of training) and complete the After Learning column.
| Gang Member | Before Learning Map Node (1-5) |
After Learning Map Node (1-5) |
|---|---|---|
| Robin | ||
| Bill | ||
| Mike | ||
| Joan | ||
| Catherine | ||
| John | ||
| Joshua | ||
| Bert | ||
| Margaret | ||
| Janet | ||
| Alfred | ||
| Gerry | ||
| Brett | ||
| Sandra | ||
| Beth | ||
| Maria |
Exercise 2: Table one shows the features for all of the gang members. Use the table to explain the pattern of results you collected in the previous exercise.
| Name | Age | Education | Marital Status | Occupation |
|---|---|---|---|---|
| Robin | 30s | Col | Single | Pusher |
| Bill | 40s | Col | Single | Pusher |
| Mike | 20s | HS | Single | Pusher |
| Joan | 20s | JH | Single | Pusher |
| Catherine | 20s | Col | Married | Pusher |
| John | 20s | Col | Divorced | Pusher |
| Joshua | 20s | Col | Single | Bookie |
| Bert | 20s | Col | Single | Burglar |
| Margaret | 30s | JH | Married | Bookie |
| Janet | 20s | JH | Married | Bookie |
| Alfred | 40s | HS | Married | Bookie |
| Gerry | 40s | Col | Married | Bookie |
| Brett | 40s | JH | Single | Bookie |
| Sandra | 40s | JH | Divorced | Bookie |
| Beth | 40s | JH | Married | Pusher |
| Maria | 40s | JH | Married | Burglar |
[Next Section: Competitive Learning]
The Connectionist Models of Cognition homepage.
