Beware the beginnings: intermediate and higher-level representations in deep neural networks are strongly affected by weight initialisation.