Learning with Resistive Switching Neural Networks

With the slowdown of Moore’s law and the intensification of memory wall as well as von-Neumann bottleneck, processing-in-memory with emerging non-volatile analog devices, such as RRAMs or memristors, is a potential solution to accelerate machine learning in hardware neural networks, which may drastically improve the energy-area efficiency. In this paper, we discuss three major types of learning, namely the supervised, reinforcement, and unsupervised learning that are implemented with various 1-transistor-1-memristor (1T1R) based neural networks.

随着摩尔定律放缓以及存储墙和冯·诺依曼瓶颈的加剧，利用新兴的非易失性模拟器件（如阻变随机存取存储器或忆阻器）进行内存处理是在硬件神经网络中加速机器学习的一种潜在解决方案，这可能会大幅提高能量 - 面积效率。在本文中，我们讨论了三种主要的学习类型，即监督学习、强化学习和无监督学习，它们是通过各种基于1晶体管 - 1忆阻器（1T1R）的神经网络实现的。