{"id":196410,"date":"2025-05-20T12:04:43","date_gmt":"2025-05-20T11:04:43","guid":{"rendered":"https:\/\/liora.io\/en\/?p=196410"},"modified":"2026-02-12T10:02:59","modified_gmt":"2026-02-12T09:02:59","slug":"all-about-self-organizing-maps","status":"publish","type":"post","link":"https:\/\/liora.io\/en\/all-about-self-organizing-maps","title":{"rendered":"Self-Organizing Maps (SOM): What are they and how to use them?"},"content":{"rendered":"\n

Self-Organizing Maps, or SOM, represent a form of artificial neural network (ANN) employed for unsupervised learning. They facilitate the reduction of data dimensionality while retaining their topological structure, thus offering a robust tool for clustering and data exploration.<\/strong><\/p>\n\n\n\n

Unlike traditional neural networks, self-organizing maps function via competitive learning<\/b> as opposed to error correction. They incorporate a neighborhood function to preserve the spatial relationships of the data<\/b>.<\/p>\n\n\n\n

\n
More about SOMs<\/a><\/div>\n<\/div>\n\n\n\n

Origin of SOMs<\/h2>\n\n\n\n

Self-organizing maps were introduced in the 1980s by the Finnish researcher Teuvo Kohonen<\/b>. This is why they are also referred to as Kohonen maps<\/b>. Inspired by biological brain mechanisms, they emulate how neurons organize and classify information, forming meaningful structures.<\/p>\n\n\n\n

How SOM Works?<\/h2>\n\n\n\n

The learning process of a Self-Organizing Map relies on multiple steps that transform complex data into an organized and readable representation<\/b>. Below is a typical, step-by-step operation of a SOM.<\/p>\n\n\n\n

1. Initialization of Weights<\/h3>\n\n\n\n

Before training starts, each neuron in the map is linked to a weight vector, which is initialized randomly. This vector shares the same dimension as the input data<\/b> and embodies the identity of each neuron prior to adjustment through the learning process.<\/p>\n\n\n\n

2. Selection of an Input Sample<\/h3>\n\n\n\n

During each iteration, an input vector is randomly selected from the training dataset<\/a>. This vector signifies a data point<\/b> that the SOM must learn to organize on the map.<\/p>\n\n\n\n

3. Identification of the Best Matching Unit<\/h3>\n\n\n\n

Once the sample is chosen, the algorithm identifies the neuron whose weights are closest to this input vector. This proximity is determined using the Euclidean distance<\/b> between the input vector and the neurons. The closest neuron is pinpointed as the BMU (Best Matching Unit)<\/b>.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
\n
More about SOMs<\/a><\/div>\n<\/div>\n\n\n\n

4. Updating the Weights of the BMU and Its Neighbors<\/h3>\n\n\n\n

After locating the BMU, the algorithm adjusts its weights to align more closely with the input vector. Neighboring neurons are also updated, albeit to a lesser extent.<\/p>\n\n\n\n

The extent of this update is influenced by two primary factors:<\/p>\n\n\n\n