Space-efficient optical computing with an integrated chip diffractive neural network
An integrated diffractive optical network that is for implementing parallel Fourier transforms, convolution operations, and application-specific optical computing with reduced footprint and energy consumption.
Optical neural networks (ONNs) that exploit photonic hardware acceleration to compute complex matrix-vector multiplication, have the advantages of ultra-high bandwidth, high calculation speed, and high parallelism over electronic counterparts. Due to the advantages of ultra-compact size, high-density integration and power-efficient, silicon photonic integrated circuits (PICs) have been emerging as a very promising candidate to build large and compact computing units for optical AI computers.
Typical silicon PIC architectures to realize chip-scale ONNs use cascades of multiple Mach-Zehnder interferometers (MZIs) and microring resonator-based wavelength division multiplexing technology. Though PICs have shown great potential in integrated optics, the space utilization (directional couplers and phase modulators), energy consumption (heaters to control phase), and the complex control circuits for reducing fluctuation of the resonance wavelength restrict the development of a large-scale programmable photonic neural network.
Our recently published paper with the topic 'optical computing with an integrated chip diffractive neural network', demonstrates a new scalable integrated diffractive neural network chip using silicon PICs, which is capable of performing the parallel Fourier transform and convolution operations. Due to the utilization of on-chip compact diffractive cells (slab waveguides), both the footprint and power consumption of the proposed architecture is reduced from quadratic scaling in the input data dimensions required for MZI-based ONN architectures to linear scaling for the IDNN. This reduction in the resource scaling from quadratic to linear will have a profound impact on the realization of large-scale silicon-photonics computing circuits with current fabrication technologies.
The parallel Fourier transform and convolution operations in the IDNN chip are furtherly applied to classification tasks with Iris flower, Handwriting digit and Fashion product datasets. The experimental results with high classification accuracy validate the functionality of our IDNN chip.
If you would like to know more about this work, please read the complete study in Nature Communications, “Space-efficient optical computing with an integrated chip diffractive neural network.”
Fig. 1 | optical integrated diffractive neural networks (IDNNs). a The multi-layer neural networks. One layer contains three main parts: optical discrete Fourier transform (ODFT) operation, amplitude/phase modulation, and optical inverse discrete Fourier transform (OIDFT) operation. A nonlinear activation function is added between two layers. b IDNN operates on complex-valued inputs using coherent light. There are two matrices based on diffractive cells and a Hadamard product operation raised by phase and amplitude modulation behind the ODFT operation. c Schematics of the experimental device. The device includes four functional parts: (1) input signal preparation; (2) implementing ODFT operation; (3) modulating amplitude/phase in the Fourier domain; (4) implementing OIDFT operation.