Testing predictions of representation cost theory with CNNs