Self-Organizing Map

Self-Organizing Map (SOM) is an unsupervised learning algorithm developed by Teuvo Kohonen that maps high-dimensional input data to a low-dimensional (typically two-dimensional) lattice of neurons in a topologically preserving way. It is a form of competitive learning in which neurons compete to represent input patterns, and the winning neuron and its neighbors adjust their weights to better match the input.

The SOM algorithm produces a discretized representation of the input space that preserves the topological relationships of the data: similar inputs activate nearby neurons on the map. This makes SOMs useful for visualization, clustering, and dimensionality reduction. The map is "self-organizing" because the global structure emerges from local interactions between neurons and inputs, without external supervision.

In the context of adaptive resonance theory and other neural network architectures, the SOM represents a distinct approach to category learning: categories are not predefined but emerge from the geometry of the data manifold, encoded in the spatial organization of the neural lattice.