Questioning Representational Optimism in Deep Learning