Even with deep learning, a common use is generating metadata. So store it, link it, and analyze it. That's what SemWeb provides. If used correctly, they compliment each other
Deep learning's cool - but it in no way competes with linked data / SemWeb, they're very different tools. Nobody seems to think Deep Learning is going to replace Postgresql, so it's definitely not going to replace linked data. Deep learning doesn't really provide you with a way to model knowledge or data, e.g. it might let you train a function to categorise photos into buckets, but it doesn't really have anything to say about the buckets and how they relate to each other. Linked Data is one way to represent that knowledge and data model.
I've never seen anyone advocate deep learning as the one way to write a whole application... You can't deep learn a rails app into existence... at some point you need humans to write code - not everything can or should be a trained/learned function.