Capacitive neural network with neuro-transistors

Zhongrui Wang, Mingyi Rao, Jin-Woo Han, Jiaming Zhang, Peng Lin, Yunning Li, Can Li, Wenhao Song, Shiva Asapu, Rivu Midya, Ye Zhuo, Hao Jiang, Jung Ho Yoon, Navnidhi K. Upadhyay, Saumil Joshi, Miao Hu, John Paul Strachan, Mark Barnell, Qing Wu, Huaqiang Wu, Qinru Qiu, R. Stanley Williams, Qiangfei Xia, J. Joshua YangNature Communications 2018 · DOI

Abstract

Experimental demonstration of resistive neural networks has been the recent focus of hardware implementation of neuromorphic computing. Capacitive neural networks, which call for novel building blocks, provide an alternative physical embodiment of neural networks featuring a lower static power and a better emulation of neural functionalities. Here, we develop neuro-transistors by integrating dynamic pseudo-memcapacitors as the gates of transistors to produce electronic analogs of the soma and axon of a neuron, with "leaky integrate-and-fire" dynamics augmented by a signal gain on the output. Paired with non-volatile pseudo-memcapacitive synapses, a Hebbian-like learning mechanism is implemented in a capacitive switching network, leading to the observed associative learning. A prototypical fully integrated capacitive neural network is built and used to classify inputs of signals.

Introduction

Neuromorphic computing aims to replicate the brain's efficiency and parallel processing capabilities in hardware. While resistive neural networks using memristors have shown promise, they often suffer from high static power consumption due to Joule heating and a lack of inherent signal gain. Biological neurons overcome these issues using capacitive membranes and active ion channels. This paper introduces the "neuro-transistor," a device that combines capacitive coupling with transistor-based amplification to create a more faithful and energy-efficient electronic analog of biological neural systems.

The Neuro-Transistor Architecture

The fundamental building block of this system is the neuro-transistor, which integrates a Dynamic Pseudo-Memcapacitor (DPM) directly onto the gate of a standard MOSFET. The DPM consists of a diffusive memristor (utilizing silver filaments) in series with a high-κ $Ta_2O_5$ capacitor.

This configuration allows the device to emulate the "leaky integrate-and-fire" (LIF) dynamics of a biological neuron. The diffusive memristor acts like a voltage-gated ion channel: at low voltages, it remains highly resistive, allowing the capacitor to integrate charge from incoming spikes. When the threshold is reached, the memristor switches to a low-resistance state, triggering a rapid voltage swing at the transistor gate—the electronic "firing" event.

Biomimetic design: Comparing biological neuron dynamics with the electronic neuro-transistor's temporal summation.

Biomimetic Learning and Dynamics

Beyond simple firing, the capacitive approach enables sophisticated neural behaviors:

  • Signal Gain: Unlike passive memristors, the transistor-based output provides active gain, allowing a single "neuron" to drive multiple subsequent layers (high fan-out).
  • Capacitive Synapses: The network utilizes Non-volatile Pseudo-Memcapacitors (NPM) as synapses. These devices retain their state, enabling the storage of weights as capacitance values rather than resistances, which significantly reduces static power dissipation.
  • Hardware Hebbian Learning: The system implements a "fire together, wire together" rule. When a neuro-transistor fires, it generates a back-propagating signal. If this coincides with a pre-synaptic spike, the resulting voltage across the synapse is sufficient to increase its capacitance, effectively "learning" the association.

Experimental Results

The researchers demonstrated the functionality of these neuro-transistors through several key experiments. They successfully observed both spatial and temporal summation, where the device integrated signals from different sources or over a period of time to reach the firing threshold.

In a classical conditioning demonstration (Pavlov’s Dog), the network learned to associate a neutral "bell" stimulus with a "food" stimulus. After training, the "bell" alone triggered the firing response. Furthermore, a prototypical $4 \times 4$ crossbar array was used to classify 2D input signals based on spiking frequencies, proving the feasibility of integrated capacitive computing.

Conclusion and Future Outlook

Capacitive neural networks offer a path toward ultra-low-power AI hardware by leveraging electrostatic energy storage and biomimetic dynamics. While challenges remain—such as the stochastic nature of silver filament formation and the complexities of large-scale fabrication—the neuro-transistor provides a powerful new tool for building brain-like hardware with inherent signal gain and high energy efficiency. Future work will focus on improving device endurance and scaling these prototypical arrays into more complex, functional systems.

Summary updated

Abstract

Experimental demonstration of resistive neural networks has been the recent focus of hardware implementation of neuromorphic computing. Capacitive neural networks, which call for novel building blocks, provide an alternative physical embodiment of neural networks featuring a lower static power and a better emulation of neural functionalities. Here, we develop neuro-transistors by integrating dynamic pseudo-memcapacitors as the gates of transistors to produce electronic analogs of the soma and axon of a neuron, with "leaky integrate-and-fire" dynamics augmented by a signal gain on the output. Paired with non-volatile pseudo-memcapacitive synapses, a Hebbian-like learning mechanism is implemented in a capacitive switching network, leading to the observed associative learning. A prototypical fully integrated capacitive neural network is built and used to classify inputs of signals.
BibTeX
@inproceedings{wang2018capacitive,
  author = {Zhongrui Wang and Mingyi Rao and Jin-Woo Han and Jiaming Zhang and Peng Lin and Yunning Li and Can Li and Wenhao Song and Shiva Asapu and Rivu Midya and Ye Zhuo and Hao Jiang and Jung Ho Yoon and Navnidhi K. Upadhyay and Saumil Joshi and Miao Hu and John Paul Strachan and Mark Barnell and Qing Wu and Huaqiang Wu and Qinru Qiu and R. Stanley Williams and Qiangfei Xia and J. Joshua Yang},
  title = {{Capacitive neural network with neuro-transistors}},
  booktitle = {Nature Communications},
  year = {2018},
  doi = {10.1038/s41467-018-05677-5},
}
Figure