This paper presents an approach to the joint optimization of neural network structure and weights which can take advantage of backpropagation as a specialized decoder. The approach is applied to binary classification of brain waves in the context of brain-computer interfaces.