Simulation of coupled spin torque oscillators for pattern recognition

Abstract

In this paper, we use circuit-level simulations to investigate the synchronization dynamics of spin torque oscillators (STOs) and demonstrate a pattern recognition scheme based on STO dynamics. We perform a sensitivity analysis in order to determine the robustness of the different STO coupling methods, considering parameter variations, such as radius or thickness of STOs. After pointing out the advantages of the cross coupled architecture, we demonstrate a coupling scheme for pattern recognition. Several patterns are encoded in the resistive coupling network and are successfully identified, after the network has been initialized correspondingly. Finally, by extracting the corresponding impulse sensitivity function, we show how an existing analytical phase domain model can be adapted to the STO array. We compare the results of this analytical model with the numerical model of the STO and conclude that they agree well providing thus a fast method for design.

Oscillatory Neural Networks Using VO2 Based Phase Encoded Logic

“…The neurons in the network are replaced by oscillators and the output is determined by the phase of each one. There are contributions of mathematical analysis with simulations using phase models for neurons (Hoppensteadt and Izhikevich, 1999;Follmann et al, 2015) as well as reporting implementations with different types of oscillators [phase-locked loops and voltagecontrolled oscillators (Hoppensteadt and Izhikevich, 2000), nonvolatile logic based on magnetic tunnel junctions (Calayir and Pileggi, 2013), micro-electro-mechanical systems and a feedback loop with transconductance amplifiers (Kumar and Mohanty, 2017), comparator and a digital circuit in Jackson et al (2018), CMOS ring oscillators (Csaba et al, 2016), STOs (Popescu et al, 2018), or VO 2 (Shukla et al, 2014;Maffezzoni et al, 2015;Corti et al, 2018)]. The implementations based on VO 2 devices exhibit potential for very low energy computation (Corti et al, 2020).…”

Energy-Efficient Ferroelectric Field-Effect Transistor-Based Oscillators for Neuromorphic System Design

“…Brain-inspired neuromorphic computing provides a strong platform to implement computationally intensive operations such as associative memory, recognition, and classification, in which traditional von-Neumann paradigms lack computational efficiency due to higher power consumption, big area, reduced accuracy, and poor parallelism [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17]. A variety of computing strategies have been demonstrated in this regard to implement such non-Boolean computing systems.…”

“…A variety of computing strategies have been demonstrated in this regard to implement such non-Boolean computing systems. Cellular Neural Networks (CNN) and Oscillatory Neural Networks (ONN) are some examples [7]. Among them, oscillatory neural networks (ONNs) earn us significant area reduction, simpler structures, fast recognition speed, and high energy efficiency [1,4,9,10,14].…”

“…Coupled-oscillator neurons have become critical parts of ONNs with the possibility of performing low-power computations and offering interesting features like synchronization dynamics. To date, modern oscillatory neurons have been physically implemented using the emerging nano-scale technologies like spin-torque (ST), memristors, and metal-insulator-transitions (MIT) oxides because of their unique properties where even the most advanced CMOS nodes are left behind [1][2][3][4][5][6][7][8][9][10][11][12][13][14]22,23]. Both ST and MIT technologies show hysteretic behaviour, synchronization capabilities, and acceptable sensitivity to image contrast compared with the CMOS-based implementation.…”

“…In [14] it is shown that scaling the number of oscillators can implement a variety of image processing applications including salience detection (two oscillators), color interpretation (three oscillators), morphological operation (five oscillators), and pattern matching (nine oscillators). The coupled spin-torque oscillators (STOs) are also considered for edge detection, pattern and spoken digit recognition in [7,8,[11][12][13]24]. Neurons are designed to operate at realistic/biologic time constants (a few milliseconds) or accelerated biological time scales (i.e.…”