Self-Organizing Maps

A Self-Organizing Map (SOM) is a type of unsupervised artificial neural network devised by the Finnish researcher Teuvo Kohonen. It was first set out in his paper «Self-Organized Formation of Topologically Correct Feature Maps» (Biological Cybernetics, 1982) and later consolidated in his monograph «Self-Organizing Maps» (Springer, 1995). Its central idea is to project high-dimensional data onto a low-dimensional grid —almost always two-dimensional— so that we can actually look at it.

What sets a SOM apart is not that it reduces dimensions but what it keeps while doing so: the approximate topology of the data. Put differently, samples that lie close together in the original space tend to activate nearby neurons on the map, so the sense of neighbourhood —of what is near and what is far— survives the trip down to two dimensions in approximate form.

What they preserve and how they are trained

Each neuron on the grid carries a weight vector of the same dimension as the input. Training is competitive: for every sample the network finds the neuron whose vector best matches it —the winning neuron, or best matching unit (BMU), usually measured with Euclidean distance. It then updates not only that winner but also its neighbourhood on the grid, weighted by a neighbourhood function (typically Gaussian) that fades with distance. As the iterations proceed, both the learning rate and the neighbourhood radius shrink, and the map moves from a coarse ordering to a fine adjustment. This twofold update —winner plus neighbours— is what yields topology preservation: adjacent neurons end up responding to similar inputs.

What they are used for

A SOM is above all a tool for visualization and exploration. By folding complex data onto a navigable surface, it lets you sense clusters, gradients and empty regions that would be invisible in the original space. That is why it has served as exploratory clustering: it does not impose a number of groups in advance but lets the map suggest where structure lies. It is best read as a first survey of the terrain, not as a final classification verdict.

Limits against t-SNE, UMAP and modern clustering

SOMs require prior choices that become part of the result. The grid, its geometry, and its number of neurons are selected before training; a small grid quantizes more samples into the same unit, while a large one spreads the evidence across more units. Topology preservation is the objective of neighborhood organization, not an exact identity between distances in the original space and the map. The configuration must therefore be evaluated on the actual data and use case rather than treating the grid's visual appearance as a guarantee. The alternatives pursue specific goals: t-SNE, by van der Maaten and Hinton (2008), places high-dimensional data in two- or three-dimensional maps and was evaluated against other visualization techniques; UMAP, by McInnes, Healy, and Melville (2018), is proposed as scalable dimension reduction that is competitive with t-SNE for visualization. Neither paper establishes a universal preference over SOM or turns visually sharp groups into a clustering verdict. A SOM retains a fixed, reusable grid; the choice depends on the goal and on which structure needs preserving, not on a closed hierarchy.