# Self-organizing map A self-organizing map (SOM) is an unsupervised artificial neural network that produces a low-dimensional representation of high-dimensional data while preserving topological structure, introduced by Teuvo Kohonen in the 1980s. A self-organizing map (SOM), also known as a self-organizing feature map (SOFM) or Kohonen map, is an unsupervised machine learning technique used to produce a low-dimensional (typically two-dimensional) representation of a higher-dimensional data set while preserving the topological structure of the data. For example, a data set with p variables measured in n observations can be represented as clusters of observations with similar values, which are then visualized as a two-dimensional map where proximal clusters have more similar values than distal ones. This makes high-dimensional data easier to visualize and analyze. The SOM is a type of [artificial neural network](https://www.wikiprompt.org/wiki/neural-network) but is trained using competitive learning rather than error-correction learning such as backpropagation with gradient descent. It was introduced by Finnish professor Teuvo Kohonen in the 1980s, building on biological models of neural systems from the 1970s and morphogenesis models dating back to Alan Turing in the 1950s. SOMs create internal representations reminiscent of the cortical homunculus, a distorted neurological map of the human body based on brain areas dedicated to sensory processing. ## Training and Mapping SOMs operate in two modes: training and mapping. Training uses an input data set (the input space) to generate a lower-dimensional representation (the map space). Mapping then classifies additional input data using the generated map. The goal of training is to represent an input space with p dimensions as a map space with n dimensions, where p > n. The map space consists of nodes or neurons arranged in a hexagonal or rectangular grid, typically two-dimensional. The number and arrangement of nodes are specified beforehand based on analysis goals. Each node is associated with a weight vector representing its position in the input space. While nodes in the map space stay fixed, training moves weight vectors toward the input data by reducing a distance metric such as Euclidean distance, without spoiling the topology induced from the map space. After training, the map classifies additional observations by finding the node with the closest weight vector to the input vector. ## Learning Algorithm The learning goal is to cause different parts of the network to respond similarly to certain input patterns, partly motivated by how sensory information is handled in separate parts of the cerebral cortex. Neuron weights are initialized either to small random values or sampled evenly from the subspace spanned by the two largest principal component eigenvectors. The latter alternative speeds learning because initial weights already approximate the final SOM weights. The network is fed many example vectors representing expected inputs, usually administered several times as iterations. Training uses competitive learning: when a training example is fed, its Euclidean distance to all weight vectors is computed, and the neuron with the most similar weight vector is the best matching unit (BMU). The BMU and nearby neurons in the SOM grid adjust their weights toward the input vector, with the magnitude of change decreasing with time and grid-distance from the BMU. The update formula for a neuron v with weight vector Wv(s) is Wv(s+1) = Wv(s) + θ(u,v,s) · α(s) · (D(t) - Wv(s)), where s is the step index, t is an index into the training sample, u is the BMU index for input vector D(t), α(s) is a monotonically decreasing learning coefficient, and θ(u,v,s) is the neighborhood function giving distance between neurons u and v at step s. The training sample can be scanned systematically, randomly drawn via bootstrap sampling, or use other methods like jackknifing. The neighborhood function depends on grid-distance between the BMU and neuron v. In simplest form, it is 1 for neurons close enough to the BMU and 0 for others, but Gaussian and Mexican-hat functions are common. The neighborhood shrinks with time: broad neighborhoods enable global self-organization, while narrow ones allow weights to converge to local estimates. The learning coefficient and neighborhood function decrease steadily or in step-wise fashion every T steps. ## Applications and Significance SOMs are used for visualization, clustering, and dimensionality reduction across fields such as data analysis, pattern recognition, and bioinformatics. They excel at preserving topological relationships, making them useful for exploratory data analysis where understanding structure is key. Unlike supervised methods, SOMs require no labeled data, making them valuable in [unsupervised learning](https://www.wikiprompt.org/wiki/machine-learning) scenarios. SOMs have been applied in diverse domains, from speech recognition and image analysis to financial data mining and medical diagnostics. Their ability to create intuitive maps of complex data has made them a staple in data visualization tools. The technique remains relevant in modern [AI](https://www.wikiprompt.org/wiki/artificial-intelligence) research, often used as a baseline for comparing more complex architectures like [deep learning](https://www.wikiprompt.org/wiki/deep-learning) models. ## Relationship to Other Neural Networks SOMs differ fundamentally from feedforward networks trained with backpropagation, such as [transformers](https://www.wikiprompt.org/wiki/transformer) used in [large language models](https://www.wikiprompt.org/wiki/large-language-model). While SOMs use competitive learning and preserve topology, transformers rely on error-correction and attention mechanisms. SOMs are also distinct from [residual networks](https://www.wikiprompt.org/wiki/residual-network) and [U-Nets](https://www.wikiprompt.org/wiki/u-net), which are supervised and use gradient-based optimization. However, SOMs share the neural network paradigm of distributed processing and adaptive weights, and they can be integrated into larger systems for preprocessing or feature extraction. ## Limitations and Extensions SOMs have limitations, including sensitivity to initialization and parameter choices, and they may not scale well to extremely high-dimensional data. Variants like growing SOMs and hierarchical SOMs address some issues by dynamically adjusting map size. Despite these challenges, SOMs remain a foundational tool in unsupervised learning, with ongoing research exploring their use in hybrid models with modern techniques. ## See Also - [Artificial neural network](https://www.wikiprompt.org/wiki/neural-network) - [Machine learning](https://www.wikiprompt.org/wiki/machine-learning) - [Deep learning](https://www.wikiprompt.org/wiki/deep-learning) - [Artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) ## External links - [Wikipedia: Self-organizing map](https://en.wikipedia.org/wiki/Self-organizing_map) --- Source: https://www.wikiprompt.org/wiki/self-organizing-map License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-09-14T21:02:52.900811+00:00