All-Photonic Phase Change Spiking Neuron: Toward Fast Neural Computing using Light

Indranil Chakraborty, Gobinda Saha, Abhronil Sengupta, Kaushik Roy

Abstract: The rapid growth of brain-inspired computing coupled with the inefficiencies in the CMOS implementations of neuromorphic systems has led to intense exploration of efficient hardware implementations of the functional units of the brain, namely, neurons and synapses. However, efforts have largely been invested in implementations in the electrical domain with potential limitations of switching speed, packing density of large integrated systems and interconnect losses. As an alternative, neuromorphic engineering in the photonic domain has recently gained attention. In this work, we demonstrate a purely photonic operation of an Integrate-and-Fire Spiking neuron, based on the phase change dynamics of Ge$_2$Sb$_2$Te$_5$ (GST) embedded on top of a microring resonator, which alleviates the energy constraints of PCMs in electrical domain. We also show that such a neuron can be potentially integrated with on-chip synapses into an all-Photonic Spiking Neural network inferencing framework which promises to be ultrafast and can potentially offer a large operating bandwidth.

Introduction

Neuromorphic (brain-inspired) computing promises extreme efficiency, but current hardware implementations face significant bottlenecks. CMOS-based systems are often limited by interconnect losses, packing density, and switching speeds. While phase-change materials (PCMs) like GST are excellent candidates for mimicking neurons due to their non-volatile state retention, their operation in the electrical domain is relatively slow and energy-intensive. To outperform CMOS, electrical PCMs require substantial reductions in current or pulse duration. This research addresses these shortcomings by moving the entire operation into the photonic domain, leveraging the massive bandwidth of light and the ultrafast response of GST to optical pulses.

Device Architecture and Photonic Integration

The proposed design centers on a bipolar Integrate-and-Fire (LIF) neuron implemented using a GST-embedded microring resonator. The "membrane potential" of the neuron is physically represented by the crystallization state of a small GST patch placed on top of the ring.

Figure 3: (a-b) Schematics of the GST-embedded microring resonator. (c-d) Normalized transmission at the through and drop ports as a function of wavelength for different degrees of GST amorphization.

GST exhibits highly contrasting optical properties: its crystalline state is highly absorptive (low transmission), while its amorphous state is transparent (high transmission). As incident optical "spikes" interact with the device via evanescent coupling, they gradually amorphize the GST. This gradual change in the degree of amorphization provides a direct physical mapping to the mathematical integration of a biological neuron's membrane potential. This photonic approach alleviates energy constraints by avoiding the high costs of driving current through resistive materials.

Spiking Dynamics and System Operation

To handle both positive and negative synaptic weights, the architecture utilizes an integration unit with two ring resonators. Their outputs are combined using a phase modulator and an interferometer to calculate the resultant membrane potential.

Figure 6: (a) Schematic of the photonic integration and firing units. (b) Temporal evolution of the membrane potential in response to incident write pulses, showing integration toward a threshold and subsequent reset.

The "Firing Unit" consists of a photonic amplifier and a GST-on-waveguide. When the integrated signal (membrane potential) crosses a predefined threshold, the amplifier provides sufficient power to amorphize the firing unit's GST, allowing an output spike to pass through. This design supports sub-nanosecond operation, with a 1.5ns "write" cycle and a 500ps "read" cycle, significantly faster than many existing neuromorphic platforms.

Key Results and Performance

The proposed neuron was evaluated using the MNIST handwritten digit dataset. By converting a standard Artificial Neural Network (ANN) to a Spiking Neural Network (SNN) framework, the researchers tested the hardware's inferencing capabilities using a multi-physics simulation stack (COMSOL, Lumerical FDTD, and MATLAB).

  • High Accuracy: The simulated SNN achieved 98.06% accuracy on MNIST, representing only a 0.24% drop compared to an ideal software-based spiking neuron. This slight degradation is attributed to the non-linear transmission characteristics of the ring resonator.
  • Energy Efficiency: The average energy for a "write" step was approximately 4 pJ per neuron, which is substantially lower than the 14–19 pJ typically required by electrical PCM devices.
  • Wavelength Multiplexing: Because the system is photonic, different neurons can operate on different wavelengths on the same waveguide, allowing for high-density networks with minimal cross-talk.

Conclusion and Future Outlook

This work demonstrates the potential for ultrafast, energy-efficient neuromorphic computing using all-photonic components. While the current focus is on the neuron body, future research will aim to integrate these units with physical photonic synapses. Additionally, training algorithms may be modified to account for device non-linearities, and thermal management of the GST-Si-SiO2 stack will be optimized to further enhance operating speeds and stability.

Summary updated

Abstract

The rapid growth of brain-inspired computing coupled with the inefficiencies in the CMOS implementations of neuromrphic systems has led to intense exploration of efficient hardware implementations of the functional units of the brain, namely, neurons and synapses. However, efforts have largely been invested in implementations in the electrical domain with potential limitations of switching speed, packing density of large integrated systems and interconnect losses. As an alternative, neuromorphic engineering in the photonic domain has recently gained attention. In this work, we demonstrate a purely photonic operation of an Integrate-and-Fire Spiking neuron, based on the phase change dynamics of Ge$_2$Sb$_2$Te$_5$ (GST) embedded on top of a microring resonator, which alleviates the energy constraints of PCMs in electrical domain. We also show that such a neuron can be potentially integrated with on-chip synapses into an all-Photonic Spiking Neural network inferencing framework which promises to be ultrafast and can potentially offer a large operating bandwidth.
BibTeX
@misc{chakraborty2018all,
  author = {Indranil Chakraborty and Gobinda Saha and Abhronil Sengupta and Kaushik Roy},
  title = {{All-Photonic Phase Change Spiking Neuron: Toward Fast Neural Computing using Light}},
  year = {2018},
  eprint = {1804.00267},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1804.00267},
}
Figure