Toward Efficient Deep Learning With Sparse Neural Networks