Find useful edges
Individual neurons make straight cuts: left or right, above or below, inside or outside.
A deep network does not make one giant leap. It transforms the same point layer by layer, until the final cut becomes easy to make.
→ Click anywhere on the map or use the sliders. Every colored edge is a cut made by a hidden neuron.
Individual neurons make straight cuts: left or right, above or below, inside or outside.
The next layer mixes those answers. Several half-spaces become corners, wedges, and bounded areas.
The original coordinates have become a pattern of activations describing where the point belongs.
The final neuron draws one last cut in the transformed space and returns the winning class.
The network’s apparent curve is made from many small straight boundaries. Earlier layers reshape the feature space; the output layer only needs to divide that reshaped space once.
w says which hidden signals matter.
h is the point after all hidden transformations.
b positions the final decision threshold.
The network begins with raw coordinates. Each layer describes the same example in a more useful language: first edges, then regions, then shapes, then evidence for a class.
No neuron understands the whole decision. The intelligence is in the composition—simple cuts arranged so their combined boundary fits the data.