I was curious about this exact question: What are the advantages of ReLU over sigmoid function in deep neural networks?
Reading through there seems to be some confusion, mainly from user Alex R., who doesn't seem happy with many of the answers there. Although they're not happy, they've not left an answer of their own, which leaves me unsure of what I can really trust from the post. The user seems to have a fairly decent reputation, so I'd be uncomfortable with dismissing their views on the thread.
It seems that making another post "because I don't trust the last one" would be a bit off, so I'm unsure what the best thing in this scenario is.