Backpropagation

[Wheelspawn]

• Backpropagation algorithm must be correct and generalize to an arbitrary amount of layers.
• Inertia term must be added (to prevent gradient descent from adjusting the weights too quickly or slowly).
• User should be allowed to choose between stochastic learning (each weight is updated as soon as its delta is calculated) and batch learning (where all the deltas are added to a list and then iterated through the network).