Shaped Pruning for Efficient Memory Addressing in DNN Accelerators
2021 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia)(2021)
摘要
In this paper, we propose a pruning strategy called “shaped pruning” to improve the efficiency of memory access. In the proposed approach, only predefined part of a filter can have weight parameters and other parts are not allowed to have the parameters. Compared to conventional pruning schemes introducing sparse filters, through the approach, memory addressing logic can be simplified and memory a...
更多查看译文
关键词
Deep learning,Shape,Conferences,Memory management,Hardware
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要