[Paper Review] Photonic neural field on a silicon chip: large-scale, high-speed neuro-inspired computing and sensing

Satoshi Sunada, Atsushi Uchida|arXiv (Cornell University)|May 22, 2021

Neural Networks and Reservoir Computing57 references4 citations

TL;DR

This paper proposes a photonic neural field on a silicon chip that leverages spatially continuous optical interference in a multimode waveguide to enable large-scale, high-speed neuro-inspired computing. By exploiting the waveguide's high spatial degrees of freedom and nonlinear response to optical inputs, the system achieves over one peta multiply-accumulate operations per second with sub-0.2 ns memory retention, demonstrating high-accuracy chaotic time-series prediction and ultrafast optical phase sensing.

ABSTRACT

Photonic neural networks have significant potential for high-speed neural processing with low latency and ultralow energy consumption. However, the on-chip implementation of a large-scale neural network is still challenging owing to its low scalability. Herein, we propose the concept of a photonic neural field and implement it experimentally on a silicon chip to realize highly scalable neuro-inspired computing. In contrast to existing photonic neural networks, the photonic neural field is a spatially continuous field that nonlinearly responds to optical inputs, and its high spatial degrees of freedom allow for large-scale and high-density neural processing on a millimeter-scale chip. In this study, we use the on-chip photonic neural field as a reservoir of information and demonstrate a high-speed chaotic time-series prediction with low errors using a training approach similar to reservoir computing. We discuss that the photonic neural field is potentially capable of executing more than one peta multiply-accumulate operations per second for a single input wavelength on a footprint as small as a few square millimeters. In addition to processing, the photonic neural field can be used for rapidly sensing the temporal variation of an optical phase, facilitated by its high sensitivity to optical inputs. The merging of optical processing with optical sensing paves the way for an end-to-end data-driven optical sensing scheme.

Motivation & Objective

• To overcome the scalability and speed limitations of existing on-chip photonic neural networks.
• To enable large-scale, high-density neural processing on a millimeter-scale silicon chip using spatial degrees of freedom of optical fields.
• To demonstrate high-speed, low-latency neuro-inspired computing and sensing using a single photonic platform.
• To achieve peta-scale MAC operations per second with minimal energy consumption.
• To integrate optical processing and sensing into an end-to-end data-driven system.

Proposed method

• The photonic neural field is implemented using a multimode waveguide that generates a spatially continuous speckle-like interference pattern from input light.
• The system exploits the complex interference of guided modes with different phase velocities to create a high-dimensional, nonlinear response field.
• The neural field acts as a reservoir in reservoir computing, where only the readout layer is trained, enabling fast and efficient computation.
• The system's response to optical inputs is measured via intensity distribution $ I(x,t) $, which encodes temporal and spatial information.
• Memory capability is evaluated using cross-correlation between input pulses and field responses, with correlation time $ t_c \approx 0.2 $ ns.
• The number of excited guided modes is computed via numerical solution of the mode-coupling equation, confirming full excitation of modes due to diffraction coupling.

Experimental results

Research questions

• RQ1 Can a spatially continuous photonic field on a silicon chip achieve large-scale, high-speed neural processing beyond the limits of discrete-node photonic networks?
• RQ2 What is the maximum processing speed and memory retention of a photonic neural field based on multimode waveguide dynamics?
• RQ3 Can the same photonic platform simultaneously perform high-speed computing and ultrafast optical phase sensing?
• RQ4 How does the spatial degree of freedom in a multimode waveguide enable scalable neural processing with minimal energy consumption?
• RQ5 To what extent can the photonic neural field emulate complex dynamical behaviors like chaos for time-series prediction?

Key findings

• The on-chip photonic neural field achieves a processing rate exceeding one peta multiply-accumulate (MAC) operations per second for a single input wavelength on a few-square-millimeter footprint.
• The system demonstrates sub-0.2 ns memory retention, with correlation time $ t_c \approx 0.2 $ ns, as confirmed by pulse-response and memory capacity experiments.
• The memory function decays sufficiently for delay steps $ i > 3 $, indicating a memory length of 0.16–0.24 ns at 12.5 GS/s input rate.
• The neural field enables high-accuracy chaotic time-series prediction with low error, validating its utility as a reservoir computing platform.
• The system exhibits high sensitivity to optical phase variations, enabling rapid sensing of temporal phase changes without additional components.
• Numerical analysis confirms that all guided modes in the multimode waveguide are excited via diffraction coupling from a single-mode input, maximizing spatial degrees of freedom.