One good place to understand the concept of neural network in 2D classification problems that tan be visualized and understand while playing.
There are shown four 2D classification problems and yo can choose any architecture of neurons changing number of layers and size of each one.
You can change also the formula that uses the neuron to decide the result:
The results are very impressive. It’s very telling the way that hiden neurons do his own bounday work, for example to round the center points in the final decissor.
You can play in playground tensorflow