While much of the work in the design of convolutional networks over the last
five years has revolved around the empirical investigation of the importance of
depth, filter sizes, and number of feature channels, recent studies have shown
that branching, i.e., splitting the computation al