Programmable metasurfaces for future photonic artificial intelligence DOI
Loubnan Abou-Hamdan,

Emil Marinov,

Peter R. Wiecha

и другие.

Nature Reviews Physics, Год журнала: 2025, Номер unknown

Опубликована: Май 12, 2025

Язык: Английский

Two-dimensional fully ferroelectric-gated hybrid computing-in-memory hardware for high-precision and energy-efficient dynamic tracking DOI Creative Commons
Tian Lu, Junying Xue, Penghui Shen

и другие.

Science Advances, Год журнала: 2024, Номер 10(36)

Опубликована: Сен. 4, 2024

Computing in memory (CIM) breaks the conventional von Neumann bottleneck through situ processing. Monolithic integration of digital and analog CIM hardware, ensuring both high precision energy efficiency, provides a sustainable paradigm for increasingly sophisticated artificial intelligence (AI) applications but remains challenging. Here, we propose complementary metal-oxide semiconductor–compatible ferroelectric hybrid platform that consists Boolean logic triggers processing multistage cell arrays computation. The basic ferroelectric-gated units are assembled with solution-processable two-dimensional (2D) molybdenum disulfide atomic-thin channels at wafer-scale yield 96.36%, delivering on/off ratios (>10 7 ), endurance 12 long retention time years), ultralow cycle-to-cycle/device-to-device variations (~0.3%/~0.5%). Last, customize highly compact 2D system dynamic tracking, achieving accuracy 99.8% 263-fold improvement power efficiency compared to graphics units. These results demonstrate potential fully hardware developing versatile blocks AI tasks.

Язык: Английский

Процитировано

8

Programmable integrated photonic coherent matrix: Principle, configuring, and applications DOI Creative Commons
Bo Wu, Hailong Zhou, Jianji Dong

и другие.

Applied Physics Reviews, Год журнала: 2024, Номер 11(1)

Опубликована: Янв. 31, 2024

Every multi-input multi-output linear optical system can be deemed as a matrix multiplier that carries out desired transformation on the input information, such imaging, modulation, and computing. The strong programmability of has been explored proved to able bring more flexibility greater possibilities applications signal processing general digital analog Furthermore, burgeoning integrated photonics with advanced manufacturing light manipulating technology pave way for large-scale reconfigurable photonic coherent matrix. This paper reviews programmable in platform. First, theoretical basis optimizing methods three types (Mach–Zehnder interferometer mesh, multi-plane diffraction, crossbar array) are introduced. Next, we overview configuring method this their processing, neural network, logic operation, recurrent acceleration, quantum computing comprehensively reviewed. Finally, challenges opportunities discussed.

Язык: Английский

Процитировано

7

Asymmetrical estimator for training encapsulated deep photonic neural networks DOI Creative Commons
Yizhi Wang, Minjia Chen, C.C. Yao

и другие.

Nature Communications, Год журнала: 2025, Номер 16(1)

Опубликована: Март 3, 2025

Abstract Photonic neural networks (PNNs) are fast in-propagation and high bandwidth paradigms that aim to popularize reproducible NN acceleration with higher efficiency lower cost. However, the training of PNN is known be challenging, where device-to-device system-to-system variations create imperfect knowledge PNN. Despite backpropagation (BP)-based algorithms being industry standard for their robustness, generality, gradient convergence digital training, existing PNN-BP methods rely heavily on accurate intermediate state extraction or extensive computational resources deep PNNs (DPNNs). The truncated photonic signal propagation computation overhead bottleneck DPNN’s operation increase system construction Here, we introduce asymmetrical (AsyT) method, tailored encapsulated DPNNs, preserved in analogue domain entire structure. AsyT offers a lightweight solution DPNNs minimum readouts, energy-efficient operation, footprint. AsyT’s ease error tolerance, generality promote widened operational scenario despite fabrication controls. We demonstrated DPNN integrated chips, repeatably enhancing performance from in-silico BP different network structures datasets.

Язык: Английский

Процитировано

1

Measuring, processing, and generating partially coherent light with self-configuring optics DOI Creative Commons
Charles Roques‐Carmes, Shanhui Fan, David A. B. Miller

и другие.

Light Science & Applications, Год журнала: 2024, Номер 13(1)

Опубликована: Сен. 20, 2024

Язык: Английский

Процитировано

6

Physical neural networks with self-learning capabilities DOI Creative Commons
Weichao Yu, Hangwen Guo, Jiang Xiao

и другие.

Science China Physics Mechanics and Astronomy, Год журнала: 2024, Номер 67(8)

Опубликована: Июль 22, 2024

Physical neural networks are artificial that mimic synapses and neurons using physical systems or materials. These harness the distinctive characteristics of to carry out computations effectively, potentially surpassing constraints conventional digital networks. A recent advancement known as ``physical self-learning'' aims achieve learning through intrinsic processes rather than relying on external computations. This article offers a comprehensive review progress made in implementing self-learning across various systems. Prevailing strategies discussed contribute realization self-learning. Despite challenges understanding fundamental mechanism learning, this work highlights towards constructing intelligent hardware from ground up, incorporating embedded self-organizing self-adaptive dynamics

Язык: Английский

Процитировано

4

Mechanical Neural Networks with Explicit and Robust Neurons DOI Creative Commons
Mei Tie, Yuan Zhou, Chang Chen

и другие.

Advanced Science, Год журнала: 2024, Номер 11(33)

Опубликована: Июнь 19, 2024

Mechanical computing provides an information processing method to realize sensing-analyzing-actuation integrated mechanical intelligence and, when combined with neural networks, can be more efficient for data-rich cognitive tasks. The requirement of solving implicit and usually nonlinear equilibrium equations motion in training networks makes computation challenging costly. Here, explicit neuron is developed which the response directly determined without need equations. A proposed ensure robustness neuron, i.e., insensitivity defects perturbations. explicitness neurons facilitate assembly various network structures. Two exemplified a robust convolutional recurrent long short-term memory capabilities associative learning, are experimentally demonstrated. introduction streamlines design fulfilling robotic matter level intelligence.

Язык: Английский

Процитировано

3

Photonic probabilistic machine learning using quantum vacuum noise DOI Creative Commons
Seou Choi, Yannick Salamin, Charles Roques‐Carmes

и другие.

Nature Communications, Год журнала: 2024, Номер 15(1)

Опубликована: Сен. 5, 2024

Probabilistic machine learning utilizes controllable sources of randomness to encode uncertainty and enable statistical modeling. Harnessing the pure quantum vacuum noise, which stems from fluctuating electromagnetic fields, has shown promise for high speed energy-efficient stochastic photonic elements. Nevertheless, computing hardware can control these elements program probabilistic algorithms been limited. Here, we implement a computer consisting element - neuron (PPN). Our PPN is implemented in bistable optical parametric oscillator (OPO) with vacuum-level injected bias fields. We then measurement-and-feedback loop time-multiplexed PPNs electronic processors (FPGA or GPU) solve certain tasks. showcase inference image generation MNIST-handwritten digits, are representative examples discriminative generative models. In both implementations, noise used as random seed classification samples. addition, propose path towards an all-optical platform, estimated sampling rate ~1 Gbps energy consumption ~5 fJ/MAC. work paves way scalable, ultrafast, hardware.

Язык: Английский

Процитировано

3

光计算和光电智能计算研究进展 DOI

张楠 Zhang Nan,

黄郅祺 Huang Zhiqi,

张子安 Zhang Zian

и другие.

Chinese Journal of Lasers, Год журнала: 2024, Номер 51(18), С. 1800001 - 1800001

Опубликована: Янв. 1, 2024

Процитировано

3

A deep neural network model for parameter identification in deep drawing metal forming process DOI

Yingjian Guo,

Can Wang, Sutao Han

и другие.

Journal of Manufacturing Processes, Год журнала: 2024, Номер 133, С. 380 - 394

Опубликована: Ноя. 29, 2024

Язык: Английский

Процитировано

3

Annealing-inspired training of an optical neural network with ternary weights DOI Creative Commons
Anas Skalli, Mirko Goldmann,

Nasibeh Haghighi

и другие.

Communications Physics, Год журнала: 2025, Номер 8(1)

Опубликована: Фев. 13, 2025

Язык: Английский

Процитировано

0