IIUC, CNNs are just a computational trick to reduce the number of parameters in the network, and train for all possible translations at once.

So how would a network perform w.r.t. translations if it was expanded, i.e., topology similar to the CNN but with all parameters expanded, and trained on translated images?