Improving Shift Invariance in Convolutional Neural Networks with Translation Invariant Polyphase Sampling