Neuromorphic Silicon Photonic Networks (1611.02272v3)

Abstract: Photonic systems for high-performance information processing have attracted renewed interest. Neuromorphic silicon photonics has the potential to integrate processing functions that vastly exceed the capabilities of electronics. We report first observations of a recurrent silicon photonic neural network, in which connections are configured by microring weight banks. A mathematical isomorphism between the silicon photonic circuit and a continuous neural network model is demonstrated through dynamical bifurcation analysis. Exploiting this isomorphism, a simulated 24-node silicon photonic neural network is programmed using "neural compiler" to solve a differential system emulation task. A 294-fold acceleration against a conventional benchmark is predicted. We also propose and derive power consumption analysis for modulator-class neurons that, as opposed to laser-class neurons, are compatible with silicon photonic platforms. At increased scale, Neuromorphic silicon photonics could access new regimes of ultrafast information processing for radio, control, and scientific computing.

• The paper introduces a recurrent silicon photonic neural network that mimics continuous-time models via microring weight banks and bifurcation analysis.
• Simulations of a 24-node network demonstrate a 294-fold speedup over conventional CPU benchmarks in solving differential system tasks.
• The study develops power consumption models for modulator-class neurons, highlighting the potential for scalable and energy-efficient silicon photonic platforms.

Neuromorphic Silicon Photonic Networks: An Overview

The paper "Neuromorphic Silicon Photonic Networks" explores the integration of neuromorphic computing with silicon photonics to potentially surpass traditional electronic capabilities in information processing. This research focuses on developing a silicon photonic neural network that leverages analog computational models, offering insights into neuromorphic photonics within silicon-compatible platforms.

Core Contributions

The paper introduces a recurrent silicon photonic neural network configured using microring weight banks, demonstrating a mathematical isomorphism with continuous neural network models. Key highlights include:

• Dynamical Analysis: The paper details a bifurcation analysis to confirm the system's isomorphism to a continuous-time recurrent neural network (CTRNN). This involves observing dynamical behaviors such as cusp and Hopf bifurcations.
• Simulation and Performance: A simulated 24-node silicon photonic neural network, programmed using a neural compiler, achieves a 294-fold acceleration against a conventional CPU benchmark while solving a differential system emulation task.
• Power Consumption: The research introduces power consumption models for modulator-class neurons, essential for silicon photonic platform compatibility, as opposed to laser-class neurons.

Technical Insights

• Broadcast-and-Weight Protocol: The network showcases a protocol utilizing wavelength division multiplexed signals, where microring resonators perform weighting operations. The electrical outputs of these weighted sums drive nonlinear electro-optic modulation, mimicking neuron function.
• Experimental Setup: Using off-chip modulator neurons, the experimental setup validates the system’s reconfigurability and stability for broader applications.

Implications and Future Directions

This research signifies a pivotal step toward scalable neuromorphic photonic systems, capable of addressing computational bottlenecks in areas such as RF processing and scientific computing:

• Scalability: By aligning with manufacturing techniques of silicon photonics, this approach promises economic scalability, akin to integrated circuits in microelectronics.
• Theoretical and Practical Applications: The ability to analogize photonic systems with neural network models opens doors to leveraging existing theoretical frameworks and computational methods for large-scale, high-speed information processing systems.

Additionally, the integration of neuromorphic concepts with silicon photonic technologies indicates potential future developments in ultra-fast computing platforms that are energy-efficient and capable of real-time information processing.

Conclusion

The paper's findings underscore the potential of neuromorphic silicon photonic networks in transforming computational paradigms. By drawing on the strengths of both photonics and neuromorphic architectures, the research presents a promising avenue for the development of next-generation computing systems, with implications for various high-performance applications. This work lays down the groundwork for future exploration in scalable, ultrafast silicon photonic computing.