Guided by brain-like ‘spiking’ computational frameworks, neuromorphic computing—brain-inspired computing for machine intelligence—promises to realize artificial intelligence while reducing the energy requirements of computing platforms. This interdisciplinary field began with the implementation of silicon circuits for biological neural routines, but has evolved to encompass the hardware implementation of algorithms with spike-based encoding and event-driven representations. Here we provide an overview of the developments in neuromorphic computing for both algorithms and hardware and highlight the fundamentals of learning and hardware frameworks. We discuss the main challenges and the future prospects of neuromorphic computing, with emphasis on algorithm–hardware codesign. The authors review the advantages and future prospects of neuromorphic computing, a multidisciplinary engineering concept for energy-efficient artificial intelligence with brain-inspired functionality.

Towards spike-based machine intelligence with neuromorphic computing.

Novel material properties can be realized by designing waves' dispersion relations in artificial crystals. The crystal's structural length scales may range from nano- (light) up to centimeters (sound waves). Because of their emergent properties these materials are called metamaterials. Different to photonics, where the dielectric constant dominantly determines the index of refraction, in a ferromagnet the spin-wave index of refraction can be dramatically changed already by the magnetization direction. This allows a different flexibility in realizing dynamic wave guides or spin-wave switches. The present review will give an introduction into the novel functionalities of spin-wave devices, concepts for spin-wave based computing and magnonic crystals. The parameters of the magnetic metamaterials are adjusted to the spin-wave k-vector such that the magnonic band structure is designed. However, already the elementary building block of an antidot lattice, the singular hole, owns a strongly varying internal potential determined by its magnetic dipole field and a localization of spin-wave modes. Photo-magnonics reveal a way to investigate the control over the interplay between localization and delocalization of the spin-wave modes using femtosecond lasers, which is a major focus of this review. We will discuss the crucial parameters to realize free Bloch states and how, by contrast, a controlled localization might allow to gradually turn on and manipulate spin-wave interactions in spin-wave based devices in the future.

The building blocks of magnonics

The discovery of the spin-torque effect has made magnetic nanodevices realistic candidates for active elements of memory devices and applications. Magnetoresistive effects allow the read-out of increasingly small magnetic bits, and the spin torque provides an efficient tool to manipulate - precisely, rapidly and at low energy cost - the magnetic state, which is in turn the central information medium of spintronic devices. By keeping the same magnetic stack, but by tuning a device's shape and bias conditions, the spin torque can be engineered to build a variety of advanced magnetic nanodevices. Here we show that by assembling these nanodevices as building blocks with different functionalities, novel types of computing architecture can be envisaged. We focus in particular on recent concepts such as magnonics and spintronic neural networks.

Spin-torque building blocks

Neuromorphic computing has come to refer to a variety of brain-inspired computers, devices, and models that contrast the pervasive von Neumann computer architecture This biologically inspired approach has created highly connected synthetic neurons and synapses that can be used to model neuroscience theories as well as solve challenging machine learning problems The promise of the technology is to create a brain-like ability to learn and adapt, but the technical challenges are significant, starting with an accurate neuroscience model of how the brain works, to finding materials and engineering breakthroughs to build devices to support these models, to creating a programming framework so the systems can learn, to creating applications with brain-like capabilities In this work, we provide a comprehensive survey of the research and motivations for neuromorphic computing over its history We begin with a 35-year review of the motivations and drivers of neuromorphic computing, then look at the major research areas of the field, which we define as neuro-inspired models, algorithms and learning approaches, hardware and devices, supporting systems, and finally applications We conclude with a broad discussion on the major research topics that need to be addressed in the coming years to see the promise of neuromorphic computing fulfilled The goals of this work are to provide an exhaustive review of the research conducted in neuromorphic computing since the inception of the term, and to motivate further work by illuminating gaps in the field where new research is needed

A Survey of Neuromorphic Computing and Neural Networks in Hardware.

Memristors cover a gap in the capabilities of basic electronic components by remembering the history of the applied electric potentials, and are considered to bring neuromorphic computers closer by imitating the performance of synapses.[1–3] We used memristive magnetic tunnel junctions[4,5] based on MgO to demonstrate that the synaptic functionality is complemented by neuron-like behavior in these nanoscopic devices. The synaptic functionality originates in a resistance change caused by a voltage-driven oxygen vacancy motion[6] within the MgO layer. The additional functionality provided by magnetic electrodes enabled a current-driven resistance modulation due to spin-transfer torque.[7] We showed that a phenomenon known as back-hopping[8–11] leads to repeated switching between two resistance levels accompanied by current spiking, which emulates neuronal behavior. As a result, this remarkably simple system, which is composed of two magnetic layers separated by an insulator, provides a sufficient basis for the fabrication of a complete neural network. According to a simple model, the human brain consists of a network of neurons, axons and synapses. Neurons exchange electrical impulses (i.e., spikes) along their axons, and synapses are the functionalized junctions between axons and neurons. The structure of the network is defined by the spatial arrangement of neurons and axons and by the strength of each synaptic connection. In this framework, one can equate the current configuration of the network with “knowledge” and its modulation with “learning”. In the conventional (Hebbian) concept of learning,[12] synaptic strength is modified by the coincident activity of presynaptic and postsynaptic neurons. In the past few years, a series of experiments has revealed a new picture. These studies suggest that instead of mere coincidence, the precise timing of presynaptic and postsynaptic spikes (i.e., spike-timing-dependent plasticity, STDP) plays a decisive role in determining the type of synaptic modification.[13,14] This modification can lead to a persistent increase or decrease in synaptic strength, commonly referred to as long-term potentiation (LTP) or long-term depression (LTD). Software inspired by neuronal networks is widely used for specific tasks, such as pattern recognition. In addition, hardware

The memristive magnetic tunnel junction as a nanoscopic synapse-neuron system.

By using phase-resolved micro-focus Brillouin light scattering spectroscopy, we demonstrate experimentally a phenomenon of wavelength conversion of spin waves propagating in tapered Permalloy waveguides. We show that this phenomenon enables efficient excitation of spin waves with sub-micrometer wavelengths being much smaller than the width of the microstrip antenna used for the excitation. The proposed excitation mechanism removes restrictions on the spin-wave wavelength imposed by the size of the antenna and enables improvement of performances of integrated magnonic devices.

Excitation of short-wavelength spin waves in magnonic waveguides

We report on the experimental study of the spatial characteristics of high-frequency spin-wave modes localized at the edges of micrometer-size in-plane magnetized permalloy ellipses. Using a near-field Brillouin light scattering technique, we have mapped the modes with the spatial resolution of few tens of nanometers. We show that the width of the localization area strongly depends on the applied magnetic field and reduces to about 85 nm for high fields. We also demonstrate that the existing theoretical models do not appropriately describe spatial characteristics of the modes.

Mapping of localized spin-wave excitations by near-field Brillouin light scattering

MgO-based magnetic tunnel junctions with up to 230% tunnel magnetoresistance ratio at room temperature and up to 345% at 13 K are prepared. The lower electrode is either exchange-biased or free, while the top electrode is free or an exchanged-biased artificial ferrimagnet, respectively. Additionally, a pseudo-spin-valve (hard-soft switching) design with two unpinned electrodes is used. Inelastic electron-tunneling spectra for each of these systems show a strong variation in the zero-bias anomaly with a reduced peak for some of the junctions. At voltages around 200 mV additional structures are found, which are not known from junctions with lower magneto resistance, such as alumina-based junctions. We discuss the spectra for the different electrode types and compare our findings with respect to barrier material and magnetoresistance ratio.

Inelastic electron tunneling spectra of MgO-based magnetic tunnel junctions with different electrode designs

We have studied experimentally with high spatial resolution the nonlinear eigenmodes of microscopic Permalloy elliptical elements. We show that the nonlinearity affects the frequencies of the edge and the center modes in an essentially different way. This leads to repulsion of corresponding resonances and to nonlinear mode hybridization resulting in qualitative modifications of the spatial characteristics of the modes. We find that the nonlinear counterparts of the edge and the center modes simultaneously exhibit features specific for both their linear analogues.

Jana Münchenberger

Papers

The memristive magnetic tunnel junction as a nanoscopic synapse-neuron system.

Excitation of short-wavelength spin waves in magnonic waveguides

Mapping of localized spin-wave excitations by near-field Brillouin light scattering

Inelastic electron tunneling spectra of MgO-based magnetic tunnel junctions with different electrode designs

Nonlinear hybridization of the fundamental eigenmodes of microscopic ferromagnetic ellipses.