Reservoir Computing with Emergent Dynamics in a Magnetic Metamaterial

Abstract

In Materio reservoir computing (RC) leverages the response of physical systems to perform computation. Dynamic systems with emergent behaviours (where local interactions lead to complex global behaviours) are especially promising for RC, as computational capability is determined by the complexity of the transformation provided. However, it is often difficult to extract these complex behaviours via device tractable measurements that can be interfaced with standard electronics. In this paper, we measure the emergent response of interconnected magnetic nanoring arrays using simple electric transport measurements, observing distinct computationally promising dynamic behaviours in device response. Then, we employ three distinct reservoir architectures that exploit each of the behaviours to perform benchmark tasks with contrasting computational requirements on a single device with state-of-the-art accuracies for spintronic computing platforms.

Complex field reversal dynamics in nanomagnetic systems

“…An implicit choice in constructing a nanomagnetic device is whether useful information is encoded in a low-energy state or the dynamic response of the system. The prevalence of energy-based logic gates for computation has diminished and dynamic responses for reservoir computing are on the rise [8,10]. Directly driving a system overcomes some limitations of nanomagnets, allowing them to evolve at lower temperatures without freezing [7] and avoiding the critical slowing down of glassy systems stopping dynamics [44].…”

“…In particular, the field of artificial spin ice [4,5] has begun to grapple with this concept. Initially, a means of directly imaging patterned, Ising-like nanomagnets with dipolar interactions that map onto problems in statistical physics and frustrated magnetism, the field has since grown to encompass device-oriented approaches to computation and evaluate the collective behaviour of nanomagnets beyond simple, Ising spins [8–10]. Visualizing the nanomagnets in real time revealed that their fluctuations do not purely correspond to a thermal ensemble, but rather incorporate the complexities of relaxation pathways [11,12], system topology [13], deviation from ergodicity [13,14] and innate material properties [7,15].…”

Magnetization Vector Rotation Reservoir Computing Operated by Redox Mechanism

“…Reservoir computing has attracted attention as a promising way of implementing an efficient artificial intelligence that can process time series data. − This sort of computing has advantages in terms of its learning cost because of its fewer learning parameters than in conventional deep learning and its ability to mimic biological systems. , These advantages could be used to reduce the amount of information traffic since the majority of information would be preprocessed on the physical reservoir. Candidates for the physical reservoir include electrical circuits, electrochemical elements, magnetic devices, optical elements, robotic systems, ion-gating devices, and so on. − They all share three key features, i.e., nonlinear transformation, short-term memory, and the ability to map time series data to a higher dimensional space. These features are evaluated by utilizing metrics of kernel rank for the ability of the reservoir to separate different input classes, generalization rank for the ability of the reservoir to generalize similar inputs of the similar class, and memory capacity (MC) for the amount of memory in the system. ,,, In particular, the magnetic devices (i.e., spin torque oscillators, spin-wave homogeneous media, anisotropic magnetoresistance arrays, and so on) have shown high computational performance in spoken digit recognition and time series data prediction tasks, in addition to being excellent candidates for miniaturization by reaching sub-μm 2 scales. − , So far, however, physical reservoirs made from magnetic materials have needed a magnetic field and/or large electric current to be applied to them, which leads to the fatal problems of high electrical power co...…”

“…Candidates for the physical reservoir include electrical circuits, electrochemical elements, magnetic devices, optical elements, robotic systems, ion-gating devices, and so on. − They all share three key features, i.e., nonlinear transformation, short-term memory, and the ability to map time series data to a higher dimensional space. These features are evaluated by utilizing metrics of kernel rank for the ability of the reservoir to separate different input classes, generalization rank for the ability of the reservoir to generalize similar inputs of the similar class, and memory capacity (MC) for the amount of memory in the system. ,,, In particular, the magnetic devices (i.e., spin torque oscillators, spin-wave homogeneous media, anisotropic magnetoresistance arrays, and so on) have shown high computational performance in spoken digit recognition and time series data prediction tasks, in addition to being excellent candidates for miniaturization by reaching sub-μm 2 scales. − , So far, however, physical reservoirs made from magnetic materials have needed a magnetic field and/or large electric current to be applied to them, which leads to the fatal problems of high electrical power consumption and structural complexity. Thus, to reduce electric power consumption and simplify the device structure, it is necessary to find a magnetic physical reservoir that does not require a magnetic field to be applied.…”