Source-linked AI summary
Metasurface-Enabled On-Chip Multiplexed Diffractive Neural Networks in the Visible
Xuhao Luo, Yueqiang Hu, Xin Li, Xiangnian Ou, Jiajie Lai, Na Liu, Huigao Duan
TL;DR
Existing optical neural-network architectures use bulky components and do not support brain-like multitasking. This paper demonstrates a polarization-multiplexed metasurface neural network for multiple object-recognition tasks, with high neuron density and a one-hidden-layer fabricated device.
Problem
Existing optical neural-network architectures are usually implemented with bulky sources and do not support human-like multitasking.
Method
The paper develops a polarization-multiplexed metasurface-based all-optical neural network for multiple object-recognition tasks.
Results
6.25×10^6/mm^2 neuron density is achieved for a single channel, with density further boosted by multiplexing, while the network performs complex object-recognition tasks.
Takeaways & Limitations
The demonstrated platform provides an integrated all-optical approach to on-chip sensing and computing through multiplexed neural processing.
Takeaways & Limitations
The fabricated on-chip MDNN has only one hidden layer.
Abstract
from arXiv · showhide
Replacing electrons with photons is a compelling route towards light-speed, highly parallel, and low-power artificial intelligence computing. Recently, all-optical diffractive neural deep neural networks have been demonstrated. However, the existing architectures often comprise bulky components and, most critically, they cannot mimic the human brain for multitasking. Here, we demonstrate a multi-skilled diffractive neural network based on a metasurface device, which can perform on-chip multi-channel sensing and multitasking at the speed of light in the visible. The metasurface is integrated with a complementary metal oxide semiconductor imaging sensor. Polarization multiplexing scheme of the subwavelength nanostructures are applied to construct a multi-channel classifier framework for simultaneous recognition of digital and fashionable items. The areal density of the artificial neurons can reach up to 6.25x106/mm2 multiplied by the number of channels. Our platform provides an integrated solution with all-optical on-chip sensing and computing for applications in machine vision, autonomous driving, and precision medicine.
Introduction
The paper develops a visible-range, on-chip metasurface diffractive neural network for multiplexed sensing and parallel recognition, addressing the limited multitasking and bulky integration of prior optical and electronic approaches. Polarization multiplexing enables simultaneous classification across channels, with simulations and experiments validating recognition and energy focusing.
- Motivation: Existing diffractive neural networks use bulky components and conventional neural networks cannot perform multiplexed information processing.These limitations motivate a compact architecture capable of multi-skilled, parallel AI processing.
- Approach: The MDNN integrates a polarization-multiplexed metasurface with a CMOS imaging sensor for visible-range, on-chip multi-channel sensing.The framework uses metasurface subwavelength structures and supports amplitude- or phase-encoded object inputs.
- Capabilities: The classifier performs simultaneous recognition of digital and fashionable items through parallel channels and multitasking at the speed of light.The multi-channel framework is trained using computer machine learning with error backpropagation.
- Approach: Polarization-dependent meta-neurons use different effective refractive indices along crossed axes to implement multiplexing.The rectangular cross section is identified as the mechanism enabling polarization multiplexing.
Conclusion
The paper demonstrates a polarization-multiplexed metasurface diffractive neural network integrated with CMOS imaging for parallel recognition and on-chip optical processing. Its 400 nm pixels yield 6.25×10^6/mm2 neurons per channel, while experiments classify dual targets and simulations extend the architecture to five hidden layers.
- The authors theoretically designed and experimentally implemented a polarization-multiplexed metasurface-based diffractive neural network.
- The network recognizes complex objects including handwritten digits and fashion items, with the physical network integrated with CMOS imaging sensors.
- The CMOS-integrated platform is intended as a miniaturized, portable sensing-and-computing chip manufacturable through semiconductor processes.
- Metasurface multiplexing enables parallel light operations and can expand the all-optical neural network into multiple channels.
- 400 nm visible-range pixels provide an effective neuron density of 6.25×10^6/mm2 for a single channel, further boosted by multiplexing.
- The fabricated one-hidden-layer MDNN demonstrates dual-target classification, while overlay lithography is proposed for multilayer meta-neurons and a five-hidden-layer design is simulated.
- The simulated five-hidden-layer framework uses 280×280×5 meta-neurons, and pre-trained metasurfaces combined with optical sensors support parallel processing.
- The platform is presented as a basis for optical multi-skilled AI chips that may support machine-vision and related sensing applications.
Methods
The methods train multi-channel MDNNs on MNIST and Fashion-MNIST, verify them numerically, fabricate metasurfaces through semiconductor-compatible processes, and characterize outputs using optical illumination and CMOS imaging.
- Computation: The networks were implemented in Python v3.6.12 and TensorFlow v2.1.0 on a server with a GeForce RTX 2080 Ti GPU.
- Training: Multi-channel MDNNs were trained individually with cross-entropy loss to maximize signal in the target region.
- Training: Training used MNIST and Fashion-MNIST datasets with batch size 10 and learning rate 0.1.
- Verification: After training, correctness was verified using Rayleigh–Sommerfeld diffraction calculations in MATLAB.
- Fabrication: The fabricated sample was produced by metasurface fabrication followed by integration with a CMOS imaging sensor.
- Fabrication: Metasurface fabrication included deposition, overlay electron-beam lithography, lift-off, and atomic layer deposition.
- Fabrication: A TiO2 film was deposited by atomic layer deposition, then etched by ion beam etching and followed by reactive ion etching of the PMMA resist.
- Integration: The metasurface was manufactured on a Sony IMX686 CMOS chip, with a 100 μm diffraction distance between hidden layers and a 100 μm optically clear adhesive.