结构稀疏动态稀疏训练（英文版）

英文标题：Dynamic Sparse Training with Structured Sparsity中文摘要：本研究提出了一种基于 DST 方法的 N:M 结构稀疏化学习方案，其具有较高的稀疏度并支持常见的硬件加速方式，在理论和实践中都表现出较好的泛化性能与较低的推理时间。英文摘要：DST methods achieve state-of-the-art results in sparse neural networktraining, matching the generalization of dense models while enabling sparsetraining and inference. Although the resulting models are highly sparse andtheoretically cheaper to train, achieving speedups with unstructured sparsityon real-world hardware is challenging. In this work we prop