基于ReRAM的高性能混合精度神经网络研究

基于阻变存储器的存算一体设计可以有效提升神经网络的存储密度、能效与吞吐率。但现有研究主要集中于优化低精度下的神经网络性能，受限于复杂的硬件结构和电路的非理想特性，提升现有研究方案计算精度处理实际复杂目标时会急剧降低神经网络的性能和可靠性。本项目拟从电路，算法，结构三个层面协同设计来提升神经网络高精度计算的性能和可靠性。本项目拟设计高密度的存储阵列和低开销的外围电路实现高性能的神经网络电路；本项目拟研究电路中不同非理想因素对高精度计算的影响，构建非理想因素模型加入到网络训练当中，提升网络高精度计算的可靠性；本项目拟研究网络中不同层精度的变化对性能的影响，通过构建智能算法调整不同层的计算精度，提出混合精度算法和架构设计实现网络准确率和性能开销的权衡。因此，本项目拟提出基于阻变存储器的高性能混合精度神经网络设计，提升神经网络处理实际复杂目标的性能和可靠性，推进我国人工智能产业的进一步发展。

The computing in memory design based on ReRAM can effectively enhance the storage density, power efficiency, and throughput of neural networks. The existing research mainly focuses on optimizing the performance of neural network under low precision. To deal with actual complex targets, the neural network needs to improve its processing precision. However, limited by the complex hardware structure and the non-ideal characteristics of the circuit, improving the processing precision will dramatically reduce the performance and reliability of the neural network. This project intends to improve the performance and reliability of high-precision neural network from three different levels, which are circuit, algorithm, and structure. Firstly, this project intends to design high-density storage arrays and low-overhead peripheral circuits to realize high-performance neural network circuits. Secondly, this project intends to study the impact of different non-ideal factors on high-precision processing. By building corresponding non-ideal factor models to optimize network training, the reliability of neural network can be greatly improved. Thirdly, this project intends to study the impact of different precision on different layers. By building intelligent algorithms to adjust the processing precision of different layers, a hybrid-precision algorithm and architecture is proposed to trade off the accuracy and performance. Therefore, this project intends to propose the high-performance hybrid-precision neural network based on ReRAM, which can improve the performance and reliability of neural networks, promoting the further development of the Chinese artificial intelligence industry.