Describir: Empiricism without magic: transformational abstraction in deep convolutional neural networks.