High-Speed Data Writing Strategy Based on Error Weight Mapping

Convolutional Neural Networks (CNNs) perform well in the fields of image recognition, natural language processing, etc. As the device of choice for neuromorphic computing, memristor has the advantages of integrated storage and computation, non-volatility, simple structure, low power consumption, and high storage density. And it plays an important role in accelerating the computation speed of neural networks. However, there is a huge challenge when performing convolutional neural network computation based on memristor: the multi value precise writing of memristors is slow. This will lead to difficulties in achieving fast Network algorithm deployment and computation of neural networks. This paper focuses on memristor-based convolutional neural network computation and proposes a method for fast multivalued mapping of memristor through reducing the multi-value writing accuracy of the memristor moderately, compared with the traditional multi-value full precision writing strategy, which can achieve less than half of the overall data writing time. This work hold great promise for applications that rely on high writing speed.