An Efficient Hardware Architecture for DNN Training by Exploiting Triple Sparsity | IEEE Conference Publication | IEEE Xplore