In principle yes (a shallow net being a universal approximation), but the question is why to do it? learning invariance is hard, if you know a prior that certain invariance makes sense, it’s better to bake it into the architecture.