Pipelined Training with Stale Weights of Deep Convolutional Neural Networks