A scale-invariant feature map

We use a simple network which uses negative feedback of activation and simple Hebbian learning to self-organize in such a way as to produce a hierarchical classification network. By adding neighbourhood relations to its learning rule, we create a feature map which has the property of retaining the angular properties of the input data, i.e. vectors of similar directions are classified similarly regardless of their magnitude. We use neither re-normalization of weights nor data preprocessing in the network despite using competition based on maximizing the neuron's activation.