Each neuron does a straight separation, and a layer of neurons does polylinear / polyplanar / polyhyperplanar separations; which looks like polygon, polyhedron, polyhyperhedron.
Shape of Layers
The network is learning in the mode called generalisation, which generalise multiple inputs into known-as-the-same-thing, and another group of inputs as another known-as-the-same-thing. However, it is also called summarising in another sense, that the number of values to consider will shrink to fewer and fewer.
And thus, the shape of layers should be having fewer and fewer neurons near the output node.
