Metasurface-enabled on-chip multiplexed diffractive neural networks in the visible

Light Science & Applications, Journal Year: 2022, Volume and Issue: 11(1)

Published: May 27, 2022

Replacing electrons with photons is a compelling route toward high-speed, massively parallel, and low-power artificial intelligence computing. Recently, diffractive networks composed of phase surfaces were trained to perform machine learning tasks through linear optical transformations. However, the existing architectures often comprise bulky components and, most critically, they cannot mimic human brain for multitasking. Here, we demonstrate multi-skilled neural network based on metasurface device, which can on-chip multi-channel sensing multitasking in visible. The polarization multiplexing scheme subwavelength nanostructures applied construct classifier framework simultaneous recognition digital fashionable items. areal density neurons reach up 6.25 × 106 mm-2 multiplied by number channels. integrated mature complementary metal-oxide semiconductor imaging sensor, providing chip-scale architecture process information directly at physical layers energy-efficient ultra-fast image processing vision, autonomous driving, precision medicine.

Language: Английский

---

Prospects and applications of on-chip lasers

eLight, Journal Year: 2023, Volume and Issue: 3(1)

Published: Jan. 4, 2023

Integrated silicon photonics has sparked a significant ramp-up of investment in both academia and industry as scalable, power-efficient, eco-friendly solution. At the heart this platform is light source, which itself, been focus research development extensively. This paper sheds conveys our perspective on current state-of-the-art different aspects application-driven on-chip lasers. We tackle from two perspectives: device-level system-wide points view. In former, routes taken integrating lasers are explored material systems to chosen integration methodologies. Then, discussion shifted towards applications that show great prospects incorporating photonic integrated circuits (PIC) with active devices, namely, optical communications interconnects, phased array-based LiDAR, sensors for chemical biological analysis, quantum technologies, finally, computing. By leveraging myriad inherent attractive features photonics, aims inspire further PICs in, but not limited to, these substantial performance gains, green solutions, mass production.

Language: Английский

---

Microcomb-based integrated photonic processing unit

Nature Communications, Journal Year: 2023, Volume and Issue: 14(1)

Published: Jan. 5, 2023

The emergence of parallel convolution-operation technology has substantially powered the complexity and functionality optical neural networks (ONN) by harnessing dimension wavelength. However, this advanced architecture faces remarkable challenges in high-level integration on-chip operation. In work, convolution based on time-wavelength plane stretching approach is implemented a microcomb-driven chip-based photonic processing unit (PPU). To support operation unit, we develop dedicated control protocol, leading to record high weight precision 9 bits. Moreover, compact data loading speed enable preeminent photonic-core compute density over 1 trillion operations per second square millimeter (TOPS mm-2). Two proof-of-concept experiments are demonstrated, including image edge detection handwritten digit recognition, showing comparable capability compared that digital computer. Due performance great scalability, can potentially revolutionize sophisticated artificial intelligence tasks autonomous driving, video action recognition reconstruction.

Language: Английский

---

Photonic multiplexing techniques for neuromorphic computing

Nanophotonics, Journal Year: 2023, Volume and Issue: 12(5), P. 795 - 817

Published: Jan. 9, 2023

Abstract The simultaneous advances in artificial neural networks and photonic integration technologies have spurred extensive research optical computing (ONNs). potential to simultaneously exploit multiple physical dimensions of time, wavelength space give ONNs the ability achieve operations with high parallelism large-data throughput. Different multiplexing techniques based on these degrees freedom enabled large-scale interconnectivity linear functions. Here, we review recent different approaches multiplexing, present our outlook key needed further advance multiplexing/hybrid-multiplexing ONNs.

Language: Английский

---

All-analog photoelectronic chip for high-speed vision tasks

Nature, Journal Year: 2023, Volume and Issue: 623(7985), P. 48 - 57

Published: Oct. 25, 2023

Abstract Photonic computing enables faster and more energy-efficient processing of vision data 1–5 . However, experimental superiority deployable systems remains a challenge because complicated optical nonlinearities, considerable power consumption analog-to-digital converters (ADCs) for downstream digital vulnerability to noises system errors 1,6–8 Here we propose an all-analog chip combining electronic light (ACCEL). It has systemic energy efficiency 74.8 peta-operations per second watt speed 4.6 (more than 99% implemented by optics), corresponding three one order magnitude higher state-of-the-art processors, respectively. After applying diffractive as encoder feature extraction, the light-induced photocurrents are directly used further calculation in integrated analog without requirement converters, leading low latency 72 ns each frame. With joint optimizations optoelectronic adaptive training, ACCEL achieves competitive classification accuracies 85.5%, 82.0% 92.6%, respectively, Fashion-MNIST, 3-class ImageNet time-lapse video recognition task experimentally, while showing superior robustness low-light conditions (0.14 fJ μm −2 frame). can be across broad range applications such wearable devices, autonomous driving industrial inspections.

Language: Английский

---

Neuromorphic Computing Based on Wavelength-Division Multiplexing

IEEE Journal of Selected Topics in Quantum Electronics, Journal Year: 2022, Volume and Issue: 29(2: Optical Computing), P. 1 - 12

Published: Aug. 31, 2022

Optical neural networks (ONNs), or optical neuromorphic hardware accelerators, have the potential to dramatically enhance computing power and energy efficiency of mainstream electronic processors, due their ultra-large bandwidths up 10's terahertz together with analog architecture that avoids need for reading writing data back-and-forth.Different multiplexing techniques been employed demonstrate ONNs, amongst which wavelengthdivision (WDM) make sufficient use unique advantages optics in terms broad bandwidths.Here, we review recent advances WDM-based focusing on methods integrated microcombs implement ONNs.We present results human image processing using an convolution accelerator operating at 11 Tera operations per second.The open challenges limitations ONNs be addressed future applications are also discussed.

Language: Английский

---

Compact optical convolution processing unit based on multimode interference

Nature Communications, Journal Year: 2023, Volume and Issue: 14(1)

Published: May 24, 2023

Convolutional neural networks are an important category of deep learning, currently facing the limitations electrical frequency and memory access time in massive data processing. Optical computing has been demonstrated to enable significant improvements terms processing speeds energy efficiency. However, most present optical schemes hardly scalable since number elements typically increases quadratically with computational matrix size. Here, a compact on-chip convolutional unit is fabricated on low-loss silicon nitride platform demonstrate its capability for large-scale integration. Three 2 × correlated real-valued kernels made two multimode interference cells four phase shifters perform parallel convolution operations. Although interrelated, ten-class classification handwritten digits from MNIST database experimentally demonstrated. The linear scalability proposed design respect size translates into solid potential

Language: Английский

---

Polarization multiplexed diffractive computing: all-optical implementation of a group of linear transformations through a polarization-encoded diffractive network

Light Science & Applications, Journal Year: 2022, Volume and Issue: 11(1)

Published: May 25, 2022

Abstract Research on optical computing has recently attracted significant attention due to the transformative advances in machine learning. Among different approaches, diffractive networks composed of spatially-engineered transmissive surfaces have been demonstrated for all-optical statistical inference and performing arbitrary linear transformations using passive, free-space layers. Here, we introduce a polarization-multiplexed processor all-optically perform multiple, arbitrarily-selected through single network trained deep In this framework, an array pre-selected polarizers is positioned between trainable materials that are isotropic, target (complex-valued) uniquely assigned combinations input/output polarization states. The transmission layers optimized via learning error-backpropagation by thousands examples fields corresponding each one complex-valued combinations. Our results analysis reveal can successfully approximate implement group with negligible error when number features/neurons ( N ) approaches $$N_pN_iN_o$$ N p i o , where i o represent pixels at input output fields-of-view, respectively, p refers unique This find various applications polarization-based vision tasks.

Language: Английский

---

Deep learning with coherent VCSEL neural networks

Nature Photonics, Journal Year: 2023, Volume and Issue: 17(8), P. 723 - 730

Published: July 17, 2023

Language: Английский

---

Massively parallel universal linear transformations using a wavelength-multiplexed diffractive optical network

Advanced Photonics, Journal Year: 2023, Volume and Issue: 5(01)

Published: Jan. 9, 2023

We report deep learning-based design of a massively parallel broadband diffractive neural network for all-optically performing large group arbitrarily-selected, complex-valued linear transformations between an input and output field-of-view, each with N_i N_o pixels, respectively. This processor is composed N_w wavelength channels, which uniquely assigned to distinct target transformation. A set arbitrarily-selected can be individually performed through the same at different illumination wavelengths, either simultaneously or sequentially (wavelength scanning). demonstrate that such network, regardless its material dispersion, successfully approximate unique transforms negligible error when number neurons (N) in matches exceeds 2 x N_o. further spectral multiplexing capability (N_w) increased by increasing N; our numerical analyses confirm these conclusions > 180, e.g., ~2000 depending on upper bound approximation error. Massively parallel, wavelength-multiplexed networks will useful designing high-throughput intelligent machine vision systems hyperspectral processors perform statistical inference analyze objects/scenes properties.

Language: Английский

---