Search

We demonstrate the utility of machine learning algorithms for the design of Oscillatory Neural Networks (ONNs). After constructing a circuit model of the oscillators in a machine-learning-enabled simulator and performing Backpropagation through time (BPTT) for determining the coupling resistances between the ring oscillators, we show the design of associative memories and multi-layered ONN classifiers. The machine-learning-designed ONNs show superior performance compared to other design methods (such as Hebbian learning) and they also enable significant simplifications in the circuit topology. We demonstrate the design of multi-layered ONNs that show superior performance compared to single-layer ones. We argue Machine learning can unlock the true computing potential of ONNs hardware.

We present the design and experimental realization of a device that acts like a spin-wave lens i.e., it focuses spin waves to a specified location. The structure of the lens does not resemble any conventional lens design, it is a nonintuitive pattern produced by a machine learning algorithm. As a spin-wave design tool, we used our custom micromagnetic solver "SpinTorch" that has built-in automatic gradient calculation and can perform backpropagation through time for spin-wave propagation. The training itself is performed with the saturation magnetization of a YIG film as a variable parameter, with the goal to guide spin waves to a predefined location. We verified the operation of the device in the widely used mumax3 micromagnetic solver, and by experimental realization. For the experimental implementation, we developed a technique to create effective saturation-magnetization landscapes in YIG by direct focused-ion-beam irradiation. This allows us to rapidly transfer the nanoscale design patterns to the YIG medium, without patterning the material by etching. We measured the effective saturation magnetization corresponding to the FIB dose levels in advance and used this mapping to translate the designed scatterer to the required dose levels. Our demonstration serves as a proof of concept for a workflow that can be used to realize more sophisticated spin-wave devices with complex functionality, e.g., spin-wave signal processors, or neuromorphic devices., Comment: 5 pages, 4 figures

We demonstrate direct focused ion beam (FIB) writing as an enabling technology for realizing spin-wave-optics devices. It is shown that ion-beam irradiation changes the characteristics of YIG films on a submicron scale in a highly controlled way, allowing to engineer the magnonic index of refraction adapted to desired applications. This technique does not physically remove material, and allows rapid fabrication of high-quality architectures of modified magnetization in magnonic media with minimal edge damage (compared to more common techniques such as etching or milling). By experimentally showing magnonic versions of a number of optical devices (lenses, gratings, Fourier-domain processors) we envision this technology as the gateway to building magnonic computing devices that rival their optical counterparts in their complexity and computational power., Comment: 11 pages, 7 figures

We experimentally demonstrate the operation of a spin-wave Rowland spectrometer. In the proposed device geometry, spin waves are coherently excited on a diffraction grating and form an interference pattern that spatially separates spectral components of the incoming signal. The diffraction grating was created by focused-ion-beam irradiation, which was found to locally eliminate the ferrimagnetic properties of YIG, without removing the material. We found that in our experiments spin waves were created by an indirect mechanism, by exploiting nonlinear resonance between the grating and the coplanar waveguide. Our work paves the way for complex spin-wave optic devices -- chips that replicate the functionality of integrated optical devices on a chip-scale., Comment: 7 pages, 5 figures, presented at Joint European Magnetic Symposia (JEMS) 2020

We demonstrate the design of a neural network, where all neuromorphic computing functions, including signal routing and nonlinear activation are performed by spin-wave propagation and interference. Weights and interconnections of the network are realized by a magnetic field pattern that is applied on the spin-wave propagating substrate and scatters the spin waves. The interference of the scattered waves creates a mapping between the wave sources and detectors. Training the neural network is equivalent to finding the field pattern that realizes the desired input-output mapping. A custom-built micromagnetic solver, based on the Pytorch machine learning framework, is used to inverse-design the scatterer. We show that the behavior of spin waves transitions from linear to nonlinear interference at high intensities and that its computational power greatly increases in the nonlinear regime. We envision small-scale, compact and low-power neural networks that perform their entire function in the spin-wave domain.

It is an intriguing concept to use oscillators as fundamental building blocks of electronic computers. The idea is not new, but is currently subject to intense research as a part of the quest for 'beyond Moore' electronic devices. In this paper we give an engineering-minded survey of oscillator-based computing architectures, with the goal of understanding their promise and limitations for next-generation computing. We will mostly discuss non-Boolean, neurally-inspired computing concepts and put the emphasis on hardware and on circuits where the oscillators are realized from emerging, nanoscale building blocks. Despite all the promise that oscillatory computing holds, existing literature gives very few clear-cut arguments about the possible benefits of using oscillators in place of other analog nonlinear circuit elements. In this survey we will argue for finding the rationale of using oscillatory building blocks and call for benchmarking studies that compare oscillatory computing circuits to level-based (analog) implementations.

Operation of the array of coupled oscillators underlying the associative memory function is demonstrated for various interconnection schemes (cross-connect, star phase keying and star frequency keying) and various physical implementation of oscillators (van der Pol, phase-locked loop, spin torque). The speed of synchronization of oscillators and the evolution of the degree of matching is studied as a function of device parameters. The dependence of errors in association on the number of the memorized patterns and the distance between the test and the memorized pattern is determined for Palm, Furber and Hopfield association algorithms., Comment: 32 pages, 20 figures

We have previously proposed a way of using coupled quantum dots to construct digital computing elements - quantum-dot cellular automata (QCA). Here we consider a different approach to using coupled quantum-dot cells in an architecture which, rather that reproducing Boolean logic, uses a physical near-neighbor connectivity to construct an analog Cellular Neural Network (CNN)., Comment: 7 pages including 3 figures