Source-linked AI summary
Neuromorphic computing with nanoscale spintronic oscillators
Jacob Torrejon, Mathieu Riou, Flavio Abreu Araujo, Sumito Tsunegi, Guru Khalsa, Damien Querlioz, Paolo Bortolotti, Vincent Cros, Akio Fukushima, Hitoshi Kubota, Shinji Yuasa, M. D. Stiles, Julie Grollier
TL;DR
Nanoscale oscillators could enable dense, low-power neuromorphic computing, but their noise and limited stability have impeded demonstrations. This paper experimentally uses a nanoscale spintronic oscillator for reservoir computing and reports spoken-digit recognition with accuracy similar to state-of-the-art neural networks.
Problem
Neuromorphic computing with nanoscale oscillators lacks experimental proof of concept despite proposals for dense, low-power hardware.
Method
The study uses a single spintronic oscillator in transient dynamics as a reservoir, with masked inputs, recorded oscillator states, and trained linear output weights.
Results
The oscillator was evaluated on spoken-digit recognition using word success rate after training output weights on some utterances and testing on the remainder.
Takeaways & Limitations
A single nanoscale non-linear oscillator can serve as a hardware reservoir for temporal-signal classification.
Abstract
from arXiv · showhide
Neurons in the brain behave as non-linear oscillators, which develop rhythmic activity and interact to process information. Taking inspiration from this behavior to realize high density, low power neuromorphic computing will require huge numbers of nanoscale non-linear oscillators. Indeed, a simple estimation indicates that, in order to fit a hundred million oscillators organized in a two-dimensional array inside a chip the size of a thumb, their lateral dimensions must be smaller than one micrometer. However, despite multiple theoretical proposals, there is no proof of concept today of neuromorphic computing with nano-oscillators. Indeed, nanoscale devices tend to be noisy and to lack the stability required to process data in a reliable way. Here, we show experimentally that a nanoscale spintronic oscillator can achieve spoken digit recognition with accuracies similar to state of the art neural networks. We pinpoint the regime of magnetization dynamics leading to highest performance. These results, combined with the exceptional ability of these spintronic oscillators to interact together, their long lifetime, and low energy consumption, open the path to fast, parallel, on-chip computation based on networks of oscillators.
Methods · Samples
The study used nanoscale magnetic tunnel junctions fabricated from multilayer films and patterned into approximately 375 nm circular devices. Under dc current injection, their spin-torque-driven vortex cores gyrated at 250–400 MHz with reported stability.
- Samples: Magnetic tunnel junction films were prepared by ultra-high-vacuum magnetron sputtering.The multilayer included buffer/PtMn/Co71Fe29/Ru/Co60Fe20B20/Co70Fe30/MgO/Fe80B20/MgO/Ta/Ru.
- Samples: The films were annealed at 360 °C for 1 h, producing resistance-area products of approximately 3.6 Ωμm2.
- Samples: Circular magnetic tunnel junctions approximately 375 nm in diameter were patterned using Ar ion etching and electron-beam lithography.
- Samples: Under dc current injection, the vortex core steadily gyrated around the dot center.
- Samples: 250 MHz to 400 MHz was the vortex gyration-frequency range for the oscillators considered.
- Samples: Spin-torque-driven vortex dynamics were described as well-understood, well-controlled, and particularly stable.
Measurement set-up
The experiments inject preprocessed signals into a magnetic nano-oscillator to perform spoken digit recognition and sine/square classification. Sampling rates are task-specific: 200 MHz for spoken digits and 500 MHz for sine/square classification.
- Signal injection: Preprocessed input signal Vin is generated by a high-frequency arbitrary waveform generator and injected as current through the magnetic nano-oscillator.The implementation supports spoken digit recognition and sine/square classification tasks.
- Sampling rates: 200 MHz is the sampling rate for spoken digit recognition, corresponding to 20 points per interval of time θ.This setting is used for the spoken digit recognition task.
- Sampling rates: 500 MHz is the sampling rate for sine/square classification, corresponding to 50 points per interval of time θ.This setting is used for classifying sine and square signals.
Reservoir Computing: general concepts
Reservoir computing uses fixed recurrent nonlinear networks to transform inputs into a higher-dimensional space, training only external readout connections. A single oscillator can emulate this reservoir through masked transient dynamics, with filtered speech inputs applied sequentially and classified from trained averaged outputs.
- General concepts: A reservoir is a randomly recurrent network of nonlinear nodes that maps inputs to a higher-dimensional space while keeping internal connections fixed.Only connections between the reservoir and output layer are trained.
- General concepts: Efficient reservoirs require separation of dynamics for different inputs and approximation of dynamics for similar inputs to support classification.These dynamical properties enable inputs to become effectively distinguishable while preserving similarity where appropriate.
- Oscillator reservoir: A single oscillator emulates a reservoir in transient dynamics when a fast-varying input and multiplicative mask create virtual nodes interconnected through time.The mask compensates for lost parallelism and simplifies hardware implementation.
- Input processing: 65-channel spectrogram or 78-channel cochlear filtering divides each spoken digit into time intervals, producing amplitudes across frequency channels and intervals.The spectrogram uses Nτ = 24 to 67 intervals, whereas the cochlear model uses Nτ = 14 to 41.
- Readout and classification: θ ≈ 100 ns input intervals keep the oscillator transient while allowing it to respond, after which trained readout weights produce 10 averaged outputs and classification selects the maximum.Each interval is about five times shorter than the oscillator relaxation time; recognition is evaluated by word success rate.
Sine and squares classification
The sine/square classification task uses masked random sequences of periodic signals, with half reserved for training and half for classification. The network distinguishes sine from square trajectories using binary targets and linear-regression weights, while recording five oscillator states per neuron.
- Input preprocessing: 160-period sequences contain equal-period sines and squares, with the first half used for training and the second half for classification.Each period is discretized into 8 points separated by time step τ.
- Input preprocessing: Each discretized signal point is multiplied by a shared random binary mask containing +1 and -1 values.The mask is a binary vector rather than the spoken-digit task’s corresponding masking structure.
- Output and readout: Sine trajectories target 0 and square trajectories target 1, with the best network weights obtained by linear regression.This follows the previously described weight-fitting procedure for spoken digit recognition.
- Output and readout: Five points are recorded for each oscillator neuron instead of one when measuring the classification output.Additional measured states are used during post-processing to increase the number of available coefficients.