Drawbacks of the Self-Organizing Map

The primary difficulty in using the SOM, which we will return to in the next chapter, is the computational demand made by training. The time required for the network to learn suitable weights increases linearly with both the size of the dataset and the length of the weights vector, and quadratically with the dimension of a square map. Every part of every sample in the database must be compared with the corresponding weight at every network node, and this process must be repeated many times, usually for at least several thousand cycles. This is an incentive to minimize the number of nodes, but as the number of nodes needed to properly represent a dataset is usually unknown, trials may be needed to determine it, which requires multiple maps to be prepared with a consequent increase in computer time. 

Scientists, of course, do not create a map out of curiosity the SOM must have a purpose. The map created by a SOM may be used in several ways, which usually lie within one of four broad categories  

Association Retrieving a complete copy of a data item if only incomplete data can be fed in. 

Projection Crushing of high-dimensional data into two dimensions. 

Modeling Developing a simplified relationship between a high-dimensional vector and a simpler representation. 

---