Abstract: Leveraging sparsity to eliminate redundant computation and storage overhead is essential for enhancing the efficiency of on-device deep neural network (DNN) learning. However, due to a lack ...