Development of Magnetic Tunnel Junction-based Neuromorphic System for Spiking Neural Networks

Abstract

Neuromorphic studies, which are inspired by the way the human brain works with the extremely low-power biological neural networks that process signals, are related to the methods for implementing a biological unit architecture that imitates the functions of neurons and synapses, which are components of the brain, through non-biological elements. Through artificial neural networks composed of these unit architectures, we expect to be able to derive high-level computational processing such as learning, recognition, and reasoning can be derived. In order to imitate biological neural networks based on ultra-high-density parallel structures, it is necessary to propose low-power, small-area artificial neuromorphic components. Accordingly, there is an active research to construct a low power, small area artificial neural network using a nonvolatile next generation electronic device including a memristor, away from conventional artificial neurons and artificial synapse architecture implemented through analog electronic circuits. A magnetic tunnel junction (MTJ) device is a nonvolatile spin-based device in which the relative magnetization alignment direction of two magnetic materials is represented by a binary value. The device is known to have excellent scalability and retention characteristics. Also, it is a next generation electronic device that is compatible with CMOS process through process technology of MRAM (Magnetic random-access memory). This study is aimed at realizing the function of low power, small area neuromorphic element and neuromorphic system using MTJ element, which is a low power next generation electronic device with high potential. The contents of the study are as follows. First, the electrical characteristics of the in-plane MTJ at the 100 nm level were investigated by using two input sources (external magnetic field, spin transfer torque current) and the switching characteristics (TMR , Hc, Jc) of the device were verified. In addition to this, by analyzing the artificially tunable telegraphic switching characteristics due to the unstable state of MTJ free layer and the memristive switching characteristics caused by the oxygen vacancy transfer of MgO, which is an insulating layer of the MTJ, we evaluated the applicability of spike train generation and control of synaptic weights to neurons, respectively. (Ch. 2) Based on the switching characteristics of the evaluated MTJ elements, artificial neurons and artificial synapse architecture were proposed and their functionality was evaluated. Through MTJ's telegraphic switching, it is confirmed that the full range of switching probability (0 ~ 100%) of the device can be controlled with respect to the intensity of the fixed input value. Using this characteristic, along with the generation of spike train, it is confirmed that rate coding and temporal coding are applicable to neural coding method for spike train. We have proposed and implemented an MTJ-based neuron architecture through a simple circuit approach that converts the switching phenomena represented by the MTJ resistance into transmittable voltage signals. In addition, Hebbian learning rules implemented by biological synapses such as spike-timing-dependent plasticity (STDP) and spike-rate-dependent plasticity (SRDP) are applied to the multiple conductivity retention properties obtained by MTJ's memristive switching. Thus, we have verified the functional characteristics of artificial synapse devices that can be matched directly with biological synapses 1: 1. (Ch. 3) As an application method of neural imitation system, a hardware-based learning system capable of performing machine learning on pattern recognition was selected through a crossbar array circuit structure composed of proposed MTJ-based artificial neuromorphic components. Its performance was verified by simulation. In this process, we have proposed ‘train’ and ‘test’ techniques using spiking neural network structure using electrical pulse signals corresponding to biological spikes. We have developed the method for recognizing digit image patterns through offline learning and online learning algorithms Awareness. In terms of recognition rate, it is lowered by about 2% for the offline learning method and about 20% for the online learning method, compared with the existing software-based learning method (~ 90%). However, from the viewpoint of computational complexity, the proposed method of applying the hardware elements directly to the computing elements has much less computational complexity, which means the implementation of learning through the low power system. (Ch. 4) In the study of neuromorphic engineering, it is important to find suitable devices firstly. Through this study, we verified the possibility through various verification of next generation electronic device called MTJ. In addition, since the MTJ can implement both artificial neurons and synapses in the same device, the advantage of the process in actual circuit integration of the neuromorphic system can be achieved. It is expected that the contents of the researches related to the neuromorphic device and system performed through this study will be a meaningful proposal in setting the direction in neuromorphic research.