Does a Kohonen network need labeled classes?

A standard SOM learns from input patterns without class labels, although labels may later help interpret the map.

Is the map always two-dimensional?

No. Two-dimensional grids are common for visualization, but other grid dimensions and topologies can be used.

Why update neighbors of the winner?

Coordinated neighborhood adaptation encourages nearby map units to represent related input patterns rather than acting as an unordered collection of prototypes.

AI Glossary / Deep learning / Intermediate

Kohonen Network

Definition

A Kohonen network commonly refers to a self-organizing map, or SOM: an unsupervised neural model that learns representative vectors arranged on a predefined low-dimensional grid. During learning, an input is matched to a winning prototype, and the winner and its grid neighbors are adjusted toward that input. This neighborhood-based adaptation encourages an organized map of similarities in the original feature space.

What it means

The map offers a way to explore observations described by many variables without assigning class labels in advance. Each map unit represents a prototype with the same feature dimensions as an input. The grid coordinates locate prototypes on the display, while the prototype values remain in the original feature space. These are different spaces. Nearby units are encouraged to represent similar patterns, but the map cannot generally preserve every distance or relationship perfectly when compressing complex data into a small grid.

How it works

An online SOM starts with prototype vectors and repeatedly selects a best-matching unit, often by Euclidean distance. A common update is w(j) new = w(j) old plus learning rate times neighborhood influence times [x minus w(j) old]. The neighborhood influence depends on distance from the winning unit on the grid, not simply distance between input vectors. Learning-rate and neighborhood schedules usually change during training. Batch variants update prototypes differently. Scaling, initialization, grid size, distance metric, and training schedule affect the resulting organization.

Three practical examples

Customer exploration: organize behavioral feature vectors on a map to inspect patterns without assuming the map units are validated customer categories.
Process monitoring: compare current operating measurements with learned prototypes and examine regions associated with different operating conditions.
Scientific visualization: display relationships among high-dimensional observations while inspecting original variables to interpret the map's structure.

Think of it this way

Think of arranging sample cards on a board so that similar cards tend to sit nearby, while adjusting neighboring positions together. The analogy captures organized similarity. Its limit is that a SOM learns numerical prototypes rather than moving literal records, and visual neighbors can hide distortions or uncertainty in the original data.

Worked example

Consider three prototypes on a one-dimensional grid ordered A-B-C. Their feature vectors are A = [0, 0], B = [2, 0], and C = [4, 0]. For input x = [1, 1], squared distances are 2 to A, 2 to B, and 10 to C. Suppose the implementation's stated tie rule chooses A. Use learning rate 0.5, neighborhood influence 1 for A, 0.5 for B, and 0 for C.
A moves to [0, 0] plus 0.5 times [1, 1], giving [0.5, 0.5]. B moves to [2, 0] plus 0.25 times [-1, 1], giving [1.75, 0.25]. C remains [4, 0]. B changes even though it did not win because it lies in the winner's grid neighborhood. If only the winning prototype changed, the update would resemble a different competitive vector-quantization procedure. This single step does not establish a well-organized map; it simply demonstrates the neighborhood mechanism and the importance of an explicit tie rule.

Illustrative case study

A fictional retail team trains a SOM on customer spending and purchase frequency. The map appears to separate customers mainly by spending, which was measured on a much larger numerical scale. After reviewing preprocessing and the business meaning of each variable, the team compares scaled and unscaled maps and checks the patterns against original records. It avoids naming map regions as fixed customer personalities or treating an attractive visualization as proof of useful segmentation.

Common misconceptions and mistakes

A map unit is not automatically a class label, and the number of units is not necessarily the number of natural clusters. SOMs use both feature-space distances for matching and grid-space neighborhoods for adaptation; confusing them changes the algorithm. Topology preservation is an aim rather than a guarantee of perfect geometry. Different initializations can produce rotated or otherwise different maps, and a low quantization error alone does not show that the grid preserves the relationships needed for a particular interpretation.

Business use

A Kohonen map can support exploratory analysis and prototype-based organization when users need to inspect many related measurements. Evaluate whether the map clarifies decisions better than simpler projections, clustering, or direct summaries. Document scaling and feature selection, show prototype values or feature planes, and inspect unfamiliar observations rather than forcing every point into a familiar narrative. If the map feeds a predictor or alerting system, validate that downstream use separately. Unsupervised training does not remove the need for careful evaluation.

Check your understanding

In a SOM, a nonwinning unit changes because it is adjacent to the winner on the grid. Is that an error?

Answer: No. Updating neighboring prototypes is a defining part of common SOM learning rules, with the amount controlled by the neighborhood function.

Related terms

Vector quantization
Competitive learning
Quantization error
Dimensionality reduction
Self-organizing map
best-matching unit
prototype
neighborhood function

Learn next

perform a few SOM updates on two-dimensional points and distinguish feature-space distances from grid-space distances. Then compare multiple initializations and preprocessing choices while examining both quantization quality and the map's topographic organization.

Feature space

Sources

Kohonen and Honkela, Kohonen Network, originator-authored Scholarpedia article: https://www.scholarpedia.org/article/Kohonen_networks McClelland and collaborators, Competitive Learning, original-author PDP handbook: https://web.stanford.edu/group/pdplab/pdphandbook/handbookch7.html

Previous: Knuth's Algorithm X Next: Kolmogorov - Arnold Networks