How a 16-neuron network learns to separate two spirals, shown step by step

Each hidden neuron can only draw one straight line. Training nudges all the weights 2,000 times: for the first few hundred steps almost nothing happens, then the boundary starts to bend, and by step 1,133 all 300 points are on the right side. The readouts, the boundary and the network weights all come from the actual training run. I made it with SenseNova 6.8 Code2Video, an open model that builds the whole film in code, narration timing included: https://github.com/OpenSenseNova/SenseNova-6.8-Code2Video