5 Neuromorphic Computing Using Emerging NV Memory Devices
5.1 Overview of Resistive RAMs and Ferroelectric RAMs in Neuromorphic Systems
Neural network systems, imitating the human brain, are also a type of interconnected network and use many of the same low cost, low power types of devices as other smart networks. Neural networks can require multilevel, analog like, memory devices as synapses connecting neurons. These single neuromorphic chips can be used for data processing in applications such as voice and vision recognition at the edge of the Internet of Things. Local intelligent nodes can analyze local data and send the results on providing an extra level of data security.
Various types of resistive RAMs (RRAMs) and ferroelectric RAMs (FeRAMs) have been used in artificial neural networks. These technologies can emulate synaptic plasticity and learning rules such as spike‐timing dependent plasticity (STDP) in which the synapses, which are connections between neurons, are altered in response to stimulus. The persistence of this change enables learning, that is, synaptic changes are persistent so the network state can be reactivated and stimulated further later. The multivalue property of RRAMs can be increasing or decreasing. When used as a synapse in a neural net, if the conductance is increasing, the property is called potentiation, and if the conductance is decreasing, the property is called depression.
Direct mapping can be provided using learning algorithms. The low ...