Are you referring to http://arxiv.org/pdf/1411.0247v1.pdf ? That doesn't mention sparse backward connections, but it does show that feedback weights that aren't computing the actual derivative dloss/din can till support learning. The network 'learns to learn', so to speak.