基于忆阻器的人工智能神经网络研究

Having hundreds of billions of neurons and tens of thousands of billions of synapses, our brain can work in parallel and deal with many jobs at the same time. In view of artificial intelligence, how to synthesize neuron and synapse with a single device consuming a small chip area and a low power is the key issue of integrating brain-level amount of neurons and synapses as well as verifying the brain-like functions. Memristor is a type of device that has two terminals and occupies a small chip area in nanometer. It can be realized via a connection on the chip. It also can work under a voltage below 1V consuming a low power. At the same time, the multiplying characteristics of memristor can be used to emulate the neuron functions. In this project , we will try to synthesize both neuron and synapse based on memristor. Furthermore, we will try to construct neural networks based on these neurons and synapses and verify the deep learning algorithms. Through this project, we will try to realize all-memritor neural networks, which is the cutting-edge research in the world-wide. And we will nerve some experts who are doing well in both deep learning algorithm and design of neural networks bottom-up from device level.

人脑之所以强大，在于其有近千亿个神经元和十多万亿神经突触，可以同时并行处理多个任务。从人工智能眼光来看，如何用单一器件、消耗最小的面积、最小的功耗，实现神经元和突触，是最终类脑量级神经元和突触的集成，以及能否实验验证类脑功能的关键。忆阻器是一种两端器件，在结构上与生物突触相似，尺寸很小，一个通孔连接就可以实现；它的操作电压可以很低，零点几伏特就可以使其状态改变，可以实现低功耗。同时忆阻器的积分特性可以很好地构建神经元模型。本课题计划研制出基于忆阻器的电子突触和神经元；进一步，设计并实现全忆阻器神经网络系统，验证深度学习算法，实现图像分割、运动目标识别等功能。通过本课题的研究，成功实现全忆阻器人工智能神经网络系统,使我国在全忆阻器神经网络研究处于世界前列，培养既有神经网络算法知识，又具有从器件级设计神经网络能力的专门人才。

忆阻器具有结构简单、功耗低、集成密度高的优势，有可能实现脑神经突触规模的集成。以忆阻器来实现神经网络是当前国际研究的热点。在本课题的资助下，申请人成功完成了各项任务：共发表SCI论文15篇，均标注本课题资助，在IEEE Solid State Circuit Society Singapore Chapter作2次邀请报告。具体研究成果如下：（1）基于忆阻器实现了人造神经元和突触，实现了人类神经网络中的双脉冲易化、遗忘、自我学习、联想记忆等功能；（2）基于忆阻器实现了类脑脉冲神经网络，并实现了图像识别功能；（3）验证了存内计算芯片，并实现语音识别功能，为下一步实现基于忆阻器的存内计算打下很好基础。国际合作方面邀请了聘请国际著名的忆阻器件专家新加坡南洋理工大学的Chen Tupei 教授， Norththumbria University的Fu Yongqing教授、日本群马大学的Yin You教授来 我国或者远程交流，并与这些教授合作发表了文章。在本课题支持下培养了博士4名：刘晏辰博士 、孔德钰博士、郭睿博士和王俊杰博士；培养硕士9名。