Minimizing Computation in Convolutional Neural Networks