or: Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Convolutions

Spatial information mixing and expansion of the effective receptive field of downstream neurons can be achieved by spatial shuffling rather than 3x3 convolutions.

For each channel in a stack, the receptive field is just shifted by one pixel. Different channels move in different directions.

It can dramatically reduce the required number of params in resnet16 and squeezenet, with 1/3 the params giving similar accuracy on cifar100.