Vowel recognition with four coupled spin-torque nano-oscillators

Miguel Romera, Philippe Talatchian, Sumito Tsunegi, Flavio Abreu Araujo, Vincent Cros, Paolo Bortolotti, Juan Trastoy, Kay Yakushiji, Akio Fukushima, Hitoshi Kubota, Shinji Yuasa, Maxence Ernoult, Damir Vodenicarevic, Tifenn Hirtzlin, Nicolas Locatelli, Damien Querlioz, Julie GrollierNat 2018 · OpenReview

TL;DR

Researchers have demonstrated a novel neuromorphic computing approach using four coupled spin-torque nano-oscillators (STNOs) to recognize spoken vowels. By leveraging the inherent tunability and synchronization capabilities of STNOs, they achieved high recognition rates with a remarkably small number of trainable parameters, paving the way for energy-efficient, hardware-based neural networks.

What Problem Does This Paper Address?

Traditional artificial neural networks rely on static neuron activation functions and standard arithmetic operations. While powerful, this approach can be computationally intensive and energy-demanding, especially for hardware implementations. A promising alternative is neuroinspired computing that mimics the brain's dynamic nature, using components that exhibit oscillations and synchronization. However, implementing this in hardware faces challenges: controlling and tuning these dynamic elements (like nano-oscillators) is difficult due to noise and variability. This paper tackles the problem of achieving reliable learning and pattern recognition in a hardware neural network by demonstrating that the exceptional tunability of STNOs can overcome these control issues.

Methods

The core of this work is a hardware neural network composed of four coupled spin-torque nano-oscillators. These STNOs, fabricated as magnetic tunnel junctions, act as the "neurons." Their output is their microwave emission frequency. The "synaptic connections" are established by electrically coupling the oscillators, allowing their emitted microwave currents to influence each other.

The key innovation lies in the training process. Instead of static weights, the STNO network is trained by dynamically adjusting the DC currents supplied to each oscillator. This current directly influences the oscillator's natural frequency. The researchers developed an automatic, real-time learning algorithm that iteratively adjusts these DC currents. The algorithm presents spoken vowel inputs (encoded as microwave frequencies) and observes the resulting synchronization patterns among the STNOs. If a vowel is misclassified, the DC currents are modified to shift the synchronization patterns until the network correctly categorizes the vowels. This process leverages the STNOs' ability to "phase-lock" to external signals and to each other.

Datasets

The study uses a subset of the Hillenbrand database of American English vowels. This database contains formant frequencies (characteristic spectral peaks) for seven different vowels pronounced by 37 female speakers. The formants were pre-processed and transformed into two input frequencies ($f_A$ and $f_B$) within the operating range of the STNOs. The training dataset comprised 80% of the vowel utterances, while the remaining 20% were used for testing.

Key Insights and Ideas

  • Tunability is Key: The exceptional tunability of STNO frequencies via DC current is the critical enabler for learning. This allows the network to adapt its internal states to recognize patterns.
  • Synchronization as Computation: The paper demonstrates that complex pattern classification can be achieved through the emergent phenomenon of oscillator synchronization, rather than just static signal processing. Different synchronization configurations correspond to different recognized vowels.
  • Small Network, Big Potential: A network of just four STNOs can perform non-trivial classification, highlighting the power of dynamical systems for neuromorphic computing.
  • Hardware Efficiency: The STNO approach offers a path to highly compact and potentially energy-efficient neuromorphic hardware, as STNOs are similar to existing memory cell technologies and can be scaled down significantly.

Key Results

The trained STNO network achieved impressive recognition rates for spoken vowels:

  • Training Recognition Rate: Up to 89%
  • Testing Recognition Rate: Up to 88%

These results were achieved with a remarkably small number of trainable parameters (around 30), significantly fewer than conventional neural networks for similar tasks. The performance was robust, with the network learning rapidly within 48 training steps. Simulations with ideal, noiseless oscillators showed a maximum achievable recognition rate of 94%, indicating that the experimental results are close to theoretical limits.

Limitations and Future Work

  • Scalability: While this work demonstrates the principle with four oscillators, scaling up to larger networks for more complex tasks is the next frontier. The authors suggest that cascading synchronization states between layers could be a promising approach for deep networks.
  • Synchronization Detection: The current method uses a spectrum analyzer for synchronization detection, which is not ideal for a fully integrated system. Future work could explore more energy-efficient methods like exploiting the spin-diode effect for direct voltage detection.
  • Noise and Variability: While STNOs offer good tunability, inherent noise and variability in nanoscale devices remain a challenge for large-scale deployment. The paper's success suggests that the learning algorithm can compensate for some of these imperfections.
  • Comparison to CMOS: The paper briefly compares the energy consumption of the STNO detection circuit to CMOS equivalents, suggesting a slight advantage for the STNO approach in small systems. However, the true energy benefits are expected to emerge in larger, deep networks.

Summary updated

Abstract

In recent years, artificial neural networks have become the flagship algorithm of artificial intelligence1. In these systems, neuron activation functions are static, and computing is achieved through standard arithmetic operations. By contrast, a prominent branch of neuroinspired computing embraces the dynamical nature of the brain and proposes to endow each component of a neural network with dynamical functionality, such as oscillations, and to rely on emergent physical phenomena, such as synchronization2–6, for solving complex problems with small networks7–11. This approach is especially interesting for hardware implementations, because emerging nanoelectronic devices can provide compact and energy-efficient nonlinear auto-oscillators that mimic the periodic spiking activity of biological neurons12–16. The dynamical couplings between oscillators can then be used to mediate the synaptic communication between the artificial neurons. One challenge for using nanodevices in this way is to achieve learning, which requires fine control and tuning of their coupled oscillations17; the dynamical features of nanodevices can be difficult to control and prone to noise and variability18. Here we show that the outstanding tunability of spintronic nano-oscillators—that is, the possibility of accurately controlling their frequency across a wide range, through electrical current and magnetic field—can be used to address this challenge. We successfully train a hardware network of four spin-torque nano-oscillators to recognize spoken vowels by tuning their frequencies according to an automatic real-time learning rule. We show that the high experimental recognition rates stem from the ability of these oscillators to synchronize. Our results demonstrate that non-trivial pattern classification tasks can be achieved with small hardware neural networks by endowing them with nonlinear dynamical features such as oscillations and synchronization. A network of four spin-torque nano-oscillators can be trained in real time to recognize spoken vowels, in a simple and scalable approach that could be exploited for large-scale neural networks.
BibTeX
@inproceedings{romera2018vowel,
  author = {Miguel Romera and Philippe Talatchian and Sumito Tsunegi and Flavio Abreu Araujo and Vincent Cros and Paolo Bortolotti and Juan Trastoy and Kay Yakushiji and Akio Fukushima and Hitoshi Kubota and Shinji Yuasa and Maxence Ernoult and Damir Vodenicarevic and Tifenn Hirtzlin and Nicolas Locatelli and Damien Querlioz and Julie Grollier},
  title = {{Vowel recognition with four coupled spin-torque nano-oscillators}},
  booktitle = {Nat},
  year = {2018},
  url = {https://openreview.net/forum?id=Pc6BGv1JOPY},
}
Figure