成宇, 邢恒拓, 韩芳. 基于忆阻器阵列的卷积网络混合映射部署优化[J]. 微电子学与计算机, 2022, 39(5): 118-124. DOI: 10.19304/J.ISSN1000-7180.2021.1205
引用本文: 成宇, 邢恒拓, 韩芳. 基于忆阻器阵列的卷积网络混合映射部署优化[J]. 微电子学与计算机, 2022, 39(5): 118-124. DOI: 10.19304/J.ISSN1000-7180.2021.1205
CHENG Yu, XING Hengtuo, HAN Fan. Mixed-mapping optimization strategy for deploying convolutional network based on memristor arrays[J]. Microelectronics & Computer, 2022, 39(5): 118-124. DOI: 10.19304/J.ISSN1000-7180.2021.1205
Citation: CHENG Yu, XING Hengtuo, HAN Fan. Mixed-mapping optimization strategy for deploying convolutional network based on memristor arrays[J]. Microelectronics & Computer, 2022, 39(5): 118-124. DOI: 10.19304/J.ISSN1000-7180.2021.1205

基于忆阻器阵列的卷积网络混合映射部署优化

Mixed-mapping optimization strategy for deploying convolutional network based on memristor arrays

  • 摘要: 忆阻器阵列有望满足边缘智能对功耗、存储密度、计算时间等的要求,但在目前忆阻器阵列资源有限的前提下,很难部署网络模型.由于采用双忆阻器映射参数的方式,映射压缩后的神经网络模型仍需要大量硬件资源.混合映射同时使用单忆阻器和双忆阻器两种映射方式部署卷积网络,可以减少资源消耗,但人为设定混合映射部署方案具有偶然性,映射后的网络模型准确率不可控.针对以上问题, 提出了一种基于资源约束的粒子群算法区分卷积网络参数重要性,并对混合映射部署进行优化.为提高网络准确率,以忆阻器阵列同行字线映射的参数作为搜索细粒度;为保证解的合理性,同时采用网络准确率和忆阻器数量作为适应度值;为加快搜索速度,在计算适应度值前加了一个混合比例约束.此外,与其他优化算法性能进行了对比,并讨论了算法搜索复杂度.实验表明,对于4值忆阻器,优化后部署方案的准确率可比人为设定部署方案高出33%.这项工作有望为边缘智能提供一种友好可行的非冯诺依曼硬件解决方案.

     

    Abstract: The memristor array is expected to meet the requirements of edge intelligence for power consumption, storage density, and computing time. However, it is hard to map huge network models with little memristor arrays. Because of the dual memristor mapping way, it still needs a lot of hardware resources mapping the neural network model after being compressed. For this problem, the method to deploy convolutional network that by using single memristor and dual memristor simultaneously in a way of mixed mapping is a good solution, which can reduce resources. However, there is a contingency in manual setting, the accuracy of the mapped network model is uncontrollable. To solve this new problem, a resource-constrained particle swarm algorithm is proposed to distinguish the importance of convolutional network parameters, and optimize the mixed-mapping deployment. In order to get a better accuracy, the parameters which mapped on the same word line of memristor array are used as a fine granularity search unit. To ensure reasonableness of the solutions, the accuracy of network and the number of memristors are both used in the step of fitness calculation. And in order to speed up the search speed, a mixing ratio constraint is added before this step. In addition, the performance and search complexity are compared with other optimization algorithms. For 4-value memristor, the optimized assignment can get 33% higher precision than the manual setting assignment. This work is expected to provide a friendly and feasible non-von Neumann hardware solution by Edge Intelligence.

     

/

返回文章
返回