Self Oraganising Maps

In this Tutorial, we will discuss about a machine learning algorithm called Self Organising Maps. Before we go into the details of the model, I would like to clear the meanings of some of the terms that we will be using further frequently.

Supervised Learning

Let us suppose we are given a dataset in which the the input data along with the corresponding labels are already given. In such problems, we usually build a model which takes a new feature vector and outputs the most appropriate label for it and this type of learning is called Supervised Learning.

Unsupervised Learning

In this case we are only aware of the input feature vectors and we have no idea about the label corresponding to the input. In such cases, a specific set of algorithms are used which basically oraginises or maps the dataset in some k-dimensional plane in the form of clusters where each cluster represents an unique label. These types of learning algorithms are said to be Unsupervised.

SOM Introduction

SOM also known as Kohenen’s Self Organising Maps is an unsupervised learning algorithm which maps the dataset onto a plane and the plane can be any dimensional. Intuitively, it can be considered as a clustering algorithm. Now, if the dimension of the plane over the data is to be mapped is less than that of input data, we can say that SOM also contributes in reducing the dimension of data(data compression). But the best part of SOM is that despite of reducing the dimensions, it does not result in losing topological relationships of the data.

Architecture

SOM is a lattice of neurons distributed on a plane over which the data is to be mapped. Each neuron has its centre and weight vector. The dimesion of the weight vector is same as that of input data and the dimension of centre depends on the dimension of the plane over which we need to map the data. The figure given above is an example of a 2-D SOM where the neurons are distributed on a 2-Dimensional plane and thus their centre co-ordinates will also be 2-dimensional. Let us assume the input data has a dimension Nx1. Then each neuron will have a weight vector of dimension Nx1.

The second figure is an image showing exactly how a two dimensional SOM works. The left part of the figure is the original data distributed randomly over the plane. The right part is the output of SOM which has done a clustering of the entire dataset.