Deep convolutional networks are vulnerable to image translation or shift, partly due to
common down-sampling layers, eg, max-pooling and strided convolution. These operations …