Simulating Neurons with Microresonators: The Future of Optical Computing and Artificial Intelligence

In the past decade, optical microresonators have garnered attention in the research community due to their ability to emit self-pulsing light pulses and their potential applications in optical computing and Spiking Neural Networks (SNNs). Despite the promising prospects, this technology still faces numerous challenges in practical applications. In a recent study, scientists used microresonators to build a small-scale spiking neural network, opening up new possibilities as optical computing and artificial intelligence converge.

What are Microresonators?

Microresonators are small optical devices capable of trapping light within a tiny space, enhancing the interaction between light and materials. Microresonators are typically made from semiconductor materials or other special optical materials that exhibit nonlinear responses to light. These devices can use light to control effects such as two-photon absorption or high-order harmonic generation.

One of the key features of microresonators is their ability to generate self-pulsing light without requiring external control signals. These pulses can simulate the behavior of neurons in the biological brain, making them a potential tool for optical spiking neural networks (SNNs).

Nonlinear Optics and Optical Computing

Nonlinear optics is a field that studies phenomena that occur when light interacts with materials, resulting in effects that are not directly proportional to the intensity of the light.

One prominent application of nonlinear optics is optical computing. Instead of using electricity to control the on/off states of semiconductor transistors as in traditional processors, optical computers rely on the material’s different responses when exposed to varying light intensities as the computational unit.

Below is a comparison table between photonics and semiconductors in the context of processor and computation applications.

Criterion	Photonics	Semiconductors
Physical nature	Uses light (photons) to transmit and process information.	Uses electrons to transmit and process information.
Processing method	Light interacts with material via nonlinear optical effects, such as two-photon absorption, high-order harmonic emission, and wave harmonics generation.	Electrons move through semiconductor circuits under the influence of electric fields and voltage. Devices like transistors control electron flow.
Processing speed	Speed of light (300,000 km/s) – extremely fast, ideal for high-speed applications like optical transmission and computing.	Information transmission speed depends on the electron movement speed, approximately 10^7 cm/s – slower than the speed of light.
Parallel processing capability	Can process in parallel, with multiple light channels transmitting simultaneously through optical devices.	Limited parallel processing, electronic circuits usually process sequentially, although improvements can be made using techniques like multi-core and multi-threading.
Energy efficiency	More energy-efficient, especially when processing nonlinear operations, as light can propagate through optics without consuming excessive energy.	Higher energy consumption when processing complex operations, especially in high-speed electronic circuits and processors.
Operating temperature	Does not face significant thermal issues as optics can transmit information with minimal heat dissipation. However, photons can lose energy if not well-controlled.	High temperatures can affect the performance of semiconductor components, requiring cooling systems in powerful processors to minimize heat buildup.
Challenges	Difficult integration with current electronic circuits, requiring complex fabrication technologies and high production costs. Photon-based systems still need time for refinement.	Limits on speed and performance when using electronics for information processing, and energy consumption remains a major issue. Further research on new materials and fabrication technology improvements is needed.

What are Spiking Neural Networks (SNNs)?

Spiking Neural Networks (SNNs) are a type of neural network that simulate how neurons in the biological brain operate. Instead of using continuous signals, neurons in SNNs generate spikes when they receive enough stimulation. SNNs are capable of processing information over time and learning from spike patterns, offering a more natural and efficient simulation of how the brain processes information.

SNNs use time-dependent spikes to transmit information, enabling more energy-efficient processing compared to traditional neural networks (Deep Neural Networks – DNNs), while also allowing continuous learning and real-time processing.

This is where the self-pulsing feature of microresonators plays a crucial role in improving computational efficiency.

Below is a comparison table between Spiking Neural Networks (SNNs) and Deep Neural Networks (DNNs), two common types of neural networks in machine learning and artificial intelligence.

Criterion	Spiking Neural Networks (SNNs)	Deep Neural Networks (DNNs)
Operation method	Based on spikes and time-based signals. Each neuron generates a spike when stimulation exceeds a certain threshold.	Based on continuous signals and nonlinear activation functions to compute the output of neurons through layers.
Signal type	Time-based signals: Neurons generate spike signals over time, similar to how neurons in the biological brain operate.	Continuous signals: Neurons receive continuous signals and process them through layers to compute the output.
Learning and weight update	Based on biological learning rules such as Spike-Timing Dependent Plasticity (STDP), where weights are adjusted based on the timing and sequence of spikes.	Based on backpropagation and gradient descent to update weights according to the error between the output and target values.
Network structure	Networks can have sparse structures with strong connections between neurons, similar to biological neural networks.	Typically have dense structures with multiple hidden layers, each layer connected to the next.
Temporal processing capability	Processes signals over time: Spikes can be used to simulate learning and processing in real-time.	Fixed signal processing: Information is processed during each computation cycle (no temporal processing).
Performance and computation	Strong inreal-time data processing and continuous learning but requires specialized hardware to maximize efficiency.	Efficient for deep learning tasks, can handle complex data but consumes large computational resources and memory.
Applications	Suitable for applications requiring continuous learning, real-time processing, and brain-like learning simulations (e.g.,robotics, smart sensors, neural simulation).	Often used in deep learning applications such as image recognition, natural language processing, supervised learning, unsupervised learning.
Synaptic coupling and energy efficiency	Strong synaptic coupling between neurons (similar to the brain), which can help save energy when processing simple tasks.	Sometimes requires large computational resources and high energy consumption to train complex models, especially in large networks with many layers.
Continuous learning and retraining	Continuous learning: Can learn in real-time without needing to retrain the entire network.	Non-continuous learning: Requires retraining or fine-tuning the model when new data is available.
Brain simulation	More accurately simulates brain behavior, with spikes and biological rhythms similar to how neurons operate.	While it can simulate learning processes, it does not reflect the biological activity of the brain as accurately as SNNs.

Experiment Results Using Self-Pulsing Microresonators in Spiking Neural Networks

Previously, optical microresonators were primarily used to generate pulse signals in compact systems. However, in this experiment, microresonators were combined to form an optical neural network capable of performing learning and recognition tasks through self-pulsing light pulses.

A breakthrough in this research is that microresonators not only generate optical signals based on self-pulsing, but also:

• Store information for a certain period of time.
• Recognize the spike frequency (spike rates) of the input signals.

Now, microresonators take on a role similar to a neural neuron, opening up new possibilities for developing optical neural networks with features like long-term memory and event detection. This represents a significant step closer to simulating the behavior of neurons in the brain.

The ability to store and detect signals without relying on traditional electronic memory makes microresonators both energy-efficient and faster in optical neural networks.

Key Assumptions in This Experiment:

• Stability and scalability assumption: An important assumption is that microresonators will maintain stable performance during processing and can be scaled for larger optical neural networks.
• Continuous learning assumption: The experiment also assumes that the optical neural network using microresonators will have the ability to learn continuously and process information in real-time without retraining the entire network.

Conclusion

The experiment using self-pulsing microresonators in spiking neural networks not only opens up great potential for optical computing but also enhances performance and energy efficiency in machine learning systems. However, for this technology to be widely applicable, further research is needed to address issues related to stability, cost, and integration into large optical systems.

---

Detailed study published in Nature journal

• Biasi, S., Lugnan, A., Micheli, D. et al. Exploring the potential of self-pulsing optical microresonators for spiking neural networks and sensing. Commun Phys 7, 380 (2024). https://doi.org/10.1038/s42005-024-01869-2