Use of superordinate labels yields more robust and human-like visual representations in convolutional neural networks