Self Organising Maps — A Brief Overview
Recently, me and my teammates were tasked to work on Self Organising Maps (SOM). After working on SOM, one can easily understand why they exist and what features they provide for dimensionality reduction. SOM — to say least — is a fascinating aspect of Artificial Neural Network. Read on to get a brief overview about Self Organising Maps
Introduction
Self-Organizing Maps, also known as SOMs or Kohonen maps, are a type of artificial neural network used for unsupervised learning. They were developed by Teuvo Kohonen in the 1980s, and have since been a popular tool for data visualization and pattern recognition.
Self-Organizing Maps (SOM) have been used for almost 40 years across various application domains such as biology, geology, healthcare, industry and humanities as an interpretable tool to explore, cluster and visualize high-dimensional data sets.
The basic idea behind a self-organizing map is to create a low-dimensional representation of high-dimensional data. The network is composed of nodes, also called neurons, arranged in a two-dimensional grid. Each node has a weight vector associated with it, which determines its position in the grid.
What exactly a SOM is?
The Self-Organizing Map (SOM) is a clustering model that introduces a topological relationship between clusters. It consists in a network of two layers: an input layer, and an output layer of interconnected nodes, often called neurons or units.
Typically, the topology of this layer is chosen as a two-dimensional grid, because it can be easily visualized. This visualization capability characterizes SOM as an interpretable clustering method.
During the training process, the SOM learns to organize the input data according to the similarity of their features. This means that nodes that are close together in the grid will respond to similar input patterns. In other words, the SOM maps the input data onto a 2D grid, where neighboring nodes represent similar input patterns.
The training process is iterative, and involves three main steps: competition, collaboration and weight update. In the neighborhood competition step, the network selects the node that is closest to the input pattern, and all neighboring nodes are also activated. In the weight update/adjustment step, the weights of the activated nodes are adjusted to better match the input pattern.
So, instead of having to deal with hundreds of rows and columns (because who would want that!), the data is processed into a simplified map; that’s what we call a self-organizing map. The map provides you with a two-dimensional representation of the exact same data set; one that is easier to read.
The following example show how the nodes of map are getting affected by the cluster that they are near to and it is to be noted that only a few nodes are attracted to one cluster.
Applications
Self-organizing maps have many applications, including data visualization, clustering, and anomaly detection. They are particularly useful for visualizing high-dimensional data, such as images or text, because they can create a low-dimensional representation that preserves the structure of the data.
- SOMs have been used in many applications; most often in vision systems for vector quantization.
- Vision based autonomous robot navigation systems historically use pre-determined models of the environment.
SOMs are particularly useful for exploratory data analysis, where they help identify patterns and relationships in the data that may not be immediately obvious. They also enable researchers to visualize high-dimensional data in two or three dimensions, making it easier to interpret and communicate the results.
NOTE: A detailed overview on the working and performance of SOM will follow soon. I’ll update a link in this overview as soon as the document is ready
