Stav dette: Computation Caching for Efficient Mobile Convolutional Neural Network Inference