I wanted to watch a neural network learn [P]

I wanted to really see how a neural network learns different functions, so built an interactive demo. You can change the architecture of the network and the function it will try to approximate.

A fully-connected network with ReLU activations will create a piecewise linear function. For a network with a single layer, the maximum number of segments the network can create is 1 + the width of the layer. For instance, if you enter "3" into the hidden layers your network can have up to 4 segments.

If you add another hidden layer afterwards, it multiplies the max number of segments. So "3 3" becomes 4x4=16. After training, the network rarely achieves the maximum though.

Here is the demo: https://blog.lukesalamone.com/posts/can-a-neural-net-learn