Neuromorphic Devices and Networks Based on Memristors with Ionic Dynamics

Bio-inspired neuromorphic computing has drawn increasing attention due to its potential for massively parallel, energy efficient and fault tolerant computation, e.g. nonstructured data processing, where memristors are considered as promising building blocks for the construction of neuromorphic networks. However, the lack of clear understanding on memristive switching dynamics (especially for oxide memristors), the undesirable nonlinearity in conductance modulation, and the inherent variations in present devices have hampered the implementation of neuromorphic device and functional networks. Here this chapter presents an approach to directly resolving ion transport dynamics in oxide based memristors via electrostatic force microscopy (EFM). Utilizing this method, unambiguous evidence of oxygen ion migration, accumulation, and conduction channel formation in HfO2 has been clearly observed, providing insights into the microscopic operation principles of oxide based memristors. These understandings are subsequently used to develop novel approaches to engineering the analog switching linearity and the number of weight states in memristive synapses. Moreover, both planar and vertical multi-terminal memristive devices capable of implementing heterosynaptic plasticity are demonstrated, enriching the functionality of memristive components. After incorporation into crossbar networks, the heterosynaptic plasticity endows the devices with facilely tunable learning rate that is highly desirable for achieving optimized learning scheme with accelerated learning and high accuracy at the same time. Finally, a fuzzy restricted Boltzmann machine (FRBM) network is proposed to tolerate device variation that is intrinsic to memristors based on ionic transport mechanism, thus paving the way for highly robust neuromorphic computing based on memristors.