That's an orthogonal distinction. The methods thus far have typically been parametric, in that there's a fixed network topology and the learning algorithm adju

That's an orthogonal distinction. The methods thus far have typically been parametric, in that there's a fixed network topology and the learning algorithm adjusts the (fixed set of) weights on the edges. There's no reason, though, why you couldn't have a nonparametric version that adaptively chose the number of hidden nodes in the networks and the connectivity structure.