A Self-Organizing Map (SOM), introduced by Teuvo Kohonen in the early 1980s, is one of the oldest unsupervised neural network architectures still taught today. Its goal is simple to state: take high-dimensional data and arrange it onto a low-dimensional grid (usually 2D) such that points that are similar in the original high-dimensional space end up near each other on the grid. It is dimensionality reduction with an explicit spatial, topological structure — unlike PCA or t-SNE, the SOM's output is literally a grid of neurons, each representing a prototype region of the input space.

The architecture: a grid of prototype vectors

A SOM consists of a grid of nodes (e.g., a 10x10 grid = 100 nodes). Each node has a weight vector of the same dimensionality as the input data — effectively, each node is a prototype sitting somewhere in the original high-dimensional space. There are no hidden layers, no backpropagation, and no explicit loss function in the usual deep learning sense — the entire model is just this grid of vectors, updated through a competitive learning rule.

Training: competitive learning and neighborhood updates

For each training example, the SOM finds the single closest node by distance — the Best Matching Unit (BMU). Then it updates not just the BMU but all of its neighbors on the grid, pulling them slightly toward the input vector too, with the strength of the pull decreasing with grid distance from the BMU. This is the mechanism that creates topological organization: neighboring grid nodes are forced to represent similar regions of input space, because they get updated together whenever any of them wins.

Both the neighborhood radius and the learning rate are shrunk over training — early on, large neighborhoods create the map's rough global layout; later, small neighborhoods refine local detail.

python

Note

PCA and autoencoders reduce dimensionality but don't guarantee that nearby points in the reduced space correspond to nearby points in the original space in any structured, visualizable way. A SOM is explicitly built so that the 2D grid's geometry mirrors the data's neighborhood structure — walking across the grid corresponds to a smooth walk through regions of the input space. This is why SOMs are popular specifically as a visualization tool: you can literally look at the map and see clusters and gradual transitions.

SOMs vs. other dimensionality-reduction / clustering tools

Method	Output structure	Preserves neighborhood topology?	Needs gradient descent?
Self-Organizing Map	Fixed 2D grid of prototypes	Yes, by construction	No — competitive learning
PCA	Linear projection to k dimensions	Only approximately, linearly	No — closed-form (eigendecomposition)
t-SNE / UMAP	Non-linear 2D/3D embedding	Yes, locally, but no fixed grid	No — specialized optimization
Autoencoder	Learned latent vector	Not guaranteed	Yes
k-means	Cluster assignment + centroids	No — no spatial relationship between clusters	No — iterative but not gradient-based

Note

SOMs predate backpropagation-based deep learning and have been largely superseded for most tasks: t-SNE and UMAP produce more faithful visualizations for exploratory data analysis, and autoencoders or contrastive embeddings produce more useful learned representations for downstream tasks. SOMs still see real use in exploratory data analysis and anomaly/fault detection in industrial and scientific settings (e.g., visualizing sensor data drift, or genomic/ecological data analysis), where the explicit, interpretable grid structure is itself valuable, but they are not part of the mainstream deep learning toolkit.

What's next

SOMs show that useful representations don't need gradient descent at all. Capsule Networks return to the gradient-based world but ask a related structural question: instead of a flat grid of prototypes, can a network explicitly represent parts, wholes, and the poses relating them?

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →