Finding the Needle in the Haystack with Convolutions: on the benefits of architectural bias