QE – TE – t = 0 Each cell is a node, painted in its own color. Colors are three numbers, red, green and blue, so the map
squeezes a cube onto a sheet: similar colors end up side by side, and where the sheet can't keep a
neighborhood together, the U-matrix shows a ridge.
Trained on an image's pixels, the nodes settle where the image has many pixels of similar color: the map
becomes the image's palette. On the right, every pixel is replaced by its nearest node's color, so a map of
64 nodes repaints the image in 64 colors.
Quantization error (QE) is the mean distance from a sample to its nearest node: how well the nodes cover the data.
Topographic error (TE) is the share of samples whose nearest and second-nearest nodes are not lattice
neighbors: how often the map's order breaks, at a fold or a twist.