Self Organising Maps (Kohonen)
- A network consists of a series of nodes, each well connected to
the input features
- Each connection has a weight
- When an instance is presented (via feature values), the
nearest node is selected
- That is, the node whose input weights are closest to the feature
values wins
- In turn its weights are modified to be closer to the input and
features and
- its neighbors weights are also modified though less so
- Running the net (short term) categorises the input
- Learning is unsupervised
- Eventually, the elastic net described by the nodes describes
the space in a few dimensions (typically, 1-3)
