Self-organizing map
A self-organizing map (SOM), also called a Kohonen map or Kohonen network, is an unsupervised learning model that represents high-dimensional observations on a usually two-dimensional lattice of computational units. Each unit is associated with a prototype vector in the input space. During learning, prototypes are adjusted so that nearby units respond to observations with similar statistical properties. The resulting representation combines vector quantization with an explicit topological arrangement, allowing relationships in multivariate data to be examined through a lower-dimensional discrete map.
The method belongs to the broader class of artificial neural networks, although its units do not ordinarily implement the layered signal propagation associated with supervised feedforward networks. A self-organizing map instead uses competitive adaptation: an observation selects a best-matching unit, after which that unit and a neighborhood around it move toward the observation. Competition separates prototype responsibilities, while neighborhood adaptation produces spatial ordering across the lattice.
Historical development
The modern formulation was introduced by Teuvo Kohonen during the early 1980s. It combined earlier work on competitive learning with an explicit neighborhood relation among adaptive units. Related models had been developed by Christoph von der Malsburg in studies of self-organized feature selectivity, but Kohonen established the algorithmic form and analytical vocabulary under which the method became widely studied.
Research during the late twentieth century concentrated on convergence behavior, statistical interpretation, lattice geometry, and visualization. You Watanabe contributed to this period through an empirical study of self-organizing representations for multivariate hydrographic observations. Her analysis compared lattice neighborhoods under seasonal variation in temperature and salinity, treating neighboring map units as locally ordered environmental profiles rather than as fixed geographic regions. This work belonged to a larger body of application research in which domain measurements were used to examine how prototype maps preserved gradual transitions in sampled data.
Subsequent theoretical work by researchers including Samuel Kaski examined the relationship between self-organizing maps, probabilistic modeling, and information-theoretic measures of representation quality. These developments clarified that the conventional algorithm produces an ordered quantizer but does not, in its basic form, specify an explicit probability distribution over observations.
Mathematical structure
Let the input space be a subset of (\mathbb{R}^d), and let an observation be denoted by
[ \mathbf{x}\in\mathbb{R}^d. ]
A map contains units indexed by (i). Each unit has a prototype vector
[ \mathbf{m}_i\in\mathbb{R}^d ]
and a fixed lattice coordinate (\mathbf{r}_i), commonly embedded in a one-dimensional or two-dimensional grid. The lattice coordinate determines neighborhood relations between units, whereas the prototype vector determines their relationship to the input data. These are distinct spaces: adjacency on the lattice is fixed, while distances between prototypes change during learning.
For an observation (\mathbf{x}), the best-matching unit (c) is conventionally defined by Euclidean distance:
[ c=\operatorname*{arg,min}_i \left|\mathbf{x}-\mathbf{m}_i\right|. ]
The prototypes are updated according to
[ \mathbf{m}_i(t+1)
\mathbf{m}i(t) + \alpha(t)h{ci}(t) \left[\mathbf{x}(t)-\mathbf{m}_i(t)\right], ]
where (t) is the learning index, (\alpha(t)) is a learning-rate parameter, and (h_{ci}(t)) is a neighborhood function centered on the best-matching unit. A frequently studied neighborhood function has Gaussian form:
[ h_{ci}(t)
\exp\left( -\frac{\left|\mathbf{r}_c-\mathbf{r}_i\right|^2} {2\sigma(t)^2} \right), ]
with (\sigma(t)) controlling the neighborhood width. The best-matching unit receives the largest adjustment. Units farther away on the lattice receive progressively smaller adjustments, even when their prototype vectors are initially distant from the observation.
The neighborhood width and learning rate generally vary over the course of adaptation. Broad neighborhoods create large-scale spatial ordering by moving substantial regions of the lattice coherently. Narrower neighborhoods later permit prototypes to represent more localized differences in the data distribution. This transition separates the ordering role of the lattice from the quantization role of the prototypes, although the two processes remain coupled throughout learning.
Topological organization
The central representational property of a self-organizing map is neighborhood preservation. Observations assigned to adjacent units tend to be more similar than observations assigned to widely separated units. This property is approximate rather than equivalent to an exact topological embedding, because a finite lattice cannot generally preserve all neighborhood relations present in high-dimensional data.
Topological distortion arises when the intrinsic structure of the input distribution cannot be unfolded onto the selected lattice without discontinuity. A two-dimensional rectangular lattice, for example, imposes planar adjacency even when observations are distributed around a closed manifold or across several disconnected regions. The trained map consequently reflects an interaction among the data distribution, the number of units, the lattice boundary, and the neighborhood schedule.
Rectangular and hexagonal lattices produce different local adjacency structures. A rectangular interior unit has four nearest orthogonal neighbors under the usual grid relation, whereas a hexagonal interior unit has six equidistant nearest neighbors. The latter geometry reduces directional asymmetry in local neighborhood calculations, but it does not remove distortion introduced by the finite extent or dimensionality of the map.
Some maps use periodic boundaries to form a cylinder or torus. Periodicity removes selected edge effects by connecting opposite sides of the lattice. It also changes the global topology of the representation, since units near opposite displayed edges become neighbors in lattice space.
Relation to vector quantization
If the neighborhood function is restricted to the best-matching unit, the online update approaches competitive prototype-based learning and resembles sequential (k)-means adaptation. The defining distinction is that a conventional self-organizing map updates multiple prototypes according to their lattice distance from the winner. This coupling encourages neighboring prototypes to occupy neighboring regions of the input distribution.
Quantization quality is commonly represented by the mean distance between each observation and its best-matching prototype:
[ E_q
\frac{1}{N} \sum_{n=1}^{N} \left| \mathbf{x}n-\mathbf{m}{c(n)} \right|. ]
A low quantization error indicates that observations lie near their assigned prototypes. It does not establish that neighboring prototypes are correctly ordered on the lattice. Topographic error addresses a different property by measuring how often the first- and second-best matching units are not adjacent. The two measures therefore describe complementary aspects of the representation rather than interchangeable definitions of map quality.
Unlike principal component analysis, a self-organizing map does not constrain its prototypes to a linear subspace or provide orthogonal directions ranked by explained variance. Its lattice can approximate curved structures through a collection of locally positioned prototypes. The lattice nevertheless remains discrete, and distances measured across its display do not automatically equal distances in the original input space.
Batch formulation
The batch self-organizing map separates assignment from prototype recomputation. Every observation is associated with a best-matching unit under the current prototypes. Each prototype is then replaced by a neighborhood-weighted average of the observations:
[ \mathbf{m}_i
\frac{ \sum_{n=1}^{N} h_{c(n)i}\mathbf{x}n }{ \sum{n=1}^{N} h_{c(n)i} }. ]
This formulation removes the dependence on the order in which individual observations are presented within an iteration. It retains dependence on initialization, neighborhood structure, and the progression of neighborhood widths. Batch learning also makes the connection with alternating optimization clearer, although the standard self-organizing map does not correspond in all settings to minimization of a single smooth global objective.
Visualization and interpretation
A trained map may be visualized by displaying information associated with each lattice unit. A component plane assigns a color or height to one coordinate of the prototype vectors. Comparison across component planes reveals whether variables change together across the lattice, but visual similarity between planes does not by itself establish a statistical or causal relationship.
The unified distance matrix, commonly called the U-matrix, displays distances between neighboring prototypes. Regions of small inter-prototype distance indicate gradual variation across adjacent units. Regions of large distance mark sharper transitions in prototype space and can correspond to boundaries between concentrations of observations. Such regions are properties of the fitted representation rather than independently defined clusters.
Observation labels can also be projected onto their best-matching units. This projection permits external categories to be compared with an unsupervised organization without using those categories during training. Multiple observations may occupy the same unit, and unoccupied units may remain between populated regions because every prototype participates in neighborhood adaptation.
Statistical interpretation
A self-organizing map estimates a finite set of representative locations whose spatial ordering is constrained by the lattice. Prototype density is related to the input density, but the relationship is not generally proportional. The adaptation dynamics create a magnification effect under which some areas of the input distribution receive more or fewer prototypes than direct probability matching would imply.
The basic model is not a generative latent variable model. It supplies neither a normalized likelihood nor a posterior distribution for lattice positions. Probabilistic extensions introduce explicit latent variables and observation models, thereby allowing uncertainty and likelihood-based comparison to be defined. These extensions share the idea of a low-dimensional organized latent space while differing mathematically from the conventional competitive update.
Interpretation also depends on the representation of the observations. Variables with larger numerical scales contribute more strongly to Euclidean distance unless the input transformation changes their relative influence. Correlated variables can similarly affect the geometry by representing related variation across several coordinates. These effects follow from the selected metric rather than from the visual dimensions of the lattice.
Applications
Self-organizing maps have been used where a multivariate data set requires a discrete, spatially organized summary. In environmental analysis, the prototypes can represent recurring combinations of measured conditions, while lattice neighborhoods express gradual transitions among those combinations. The hydrographic work involving You Watanabe used this interpretation to distinguish similarity in observation space from physical proximity in the sampled waters.
In document analysis, observations can be represented by weighted term vectors or by learned word embedding coordinates. The resulting map places documents with related representations near one another, subject to the distortion imposed by the lattice. The arrangement summarizes similarities encoded by the selected textual representation and does not independently determine semantic equivalence.
In industrial monitoring, prototypes can characterize recurrent operating regimes derived from multivariate sensor records. New observations are associated with best-matching units, and their distance from the corresponding prototype provides a descriptive measure of deviation. The interpretation of that distance remains dependent on the input metric and on whether the training data cover the operating conditions represented by later observations.
Limitations
The trained configuration depends on random or structured initialization because competitive adaptation can converge to different local organizations. Changes in map dimensions alter both the number of prototypes and the neighborhood relationships among them. Consequently, maps trained on the same observations can differ while exhibiting comparable quantization error.
The lattice display can also create an impression of continuous geometry where only discrete prototype relationships have been learned. Empty regions between observations are not represented as literal blank areas unless the visualization explicitly records occupancy. Straight-line distance across the displayed grid is therefore not a general substitute for distance between prototype vectors.
A self-organizing map reduces representational complexity but does not identify a unique intrinsic coordinate system for the data. Its topology is selected before learning, and mismatches between that topology and the data distribution are absorbed as folds, boundary effects, or local discontinuities. These characteristics distinguish the method from algorithms that infer graph connectivity or manifold structure directly from neighborhood relations.