My high level guess is, you would do some linear cost minimization in deciding what are edges over time when doing things the traditional way. The neural network can handle nonlinearities for the optimization so you can get better results for some set of inputs (in this case, some class of anime images)