Category learning in deep neural networks: Information content and geometry of internal representations