Self-organising maps
Kohonen described a network that lays multidimensional data out on a two-dimensional grid so that similar inputs end up near each other.
Why it matters
Unsupervised learning produced not classes but a map: the structure of the data became visible before it was named.
Each unit holds a weight vector; for an input the nearest unit is chosen, and it and its grid neighbours move toward the input. The topology of the inputs is thereby carried onto the grid. The method was widely used for exploratory data analysis and clustering. Kohonen compared the scheme directly to topographic maps in the cortex.