Zebra: Accelerating Distributed Sparse Deep Training With in-Network Gradient Aggregation for Hot Parameters | IEEE Conference Publication | IEEE Xplore