Higher-dimensional processing using a photonic tensor core with continuous-time data

Nature Photonics, Journal Year: 2023, Volume and Issue: 17(12), P. 1080 - 1088

Published: Oct. 19, 2023

Abstract New developments in hardware-based ‘accelerators’ range from electronic tensor cores and memristor-based arrays to photonic implementations. The goal of these approaches is handle the exponentially growing computational load machine learning, which currently requires doubling hardware capability approximately every 3.5 months. One solution increasing data dimensionality that processable by such hardware. Although two-dimensional processing multiplexing space wavelength has been previously reported, use three-dimensional not yet implemented In this paper, we introduce radio-frequency modulation signals increase parallelization, adding an additional dimension alongside spatially distributed non-volatile memories multiplexing. We leverage higher-dimensional configure a system architecture compatible with edge computing frameworks. Our achieves parallelism 100, two orders higher than implementations using only spatial degrees freedom. demonstrate performing synchronous convolution 100 clinical electrocardiogram patients cardiovascular diseases, constructing convolutional neural network capable identifying at sudden death risk 93.5% accuracy.

Language: Английский

---

Optical neural networks: progress and challenges

Light Science & Applications, Journal Year: 2024, Volume and Issue: 13(1)

Published: Sept. 20, 2024

Language: Английский

---

Photonic neural networks and optics-informed deep learning fundamentals

APL Photonics, Journal Year: 2024, Volume and Issue: 9(1)

Published: Jan. 1, 2024

The recent explosive compute growth, mainly fueled by the boost of artificial intelligence (AI) and deep neural networks (DNNs), is currently instigating demand for a novel computing paradigm that can overcome insurmountable barriers imposed conventional electronic architectures. Photonic (PNNs) implemented on silicon integration platforms stand out as promising candidate to endow network (NN) hardware, offering potential energy efficient ultra-fast computations through utilization unique primitives photonics, i.e., efficiency, THz bandwidth, low-latency. Thus far, several demonstrations have revealed huge PNNs in performing both linear non-linear NN operations at unparalleled speed consumption metrics. Transforming this into tangible reality learning (DL) applications requires, however, understanding basic PNN principles, requirements, challenges across all constituent architectural, technological, training aspects. In Tutorial, we, initially, review principles DNNs along with their fundamental building blocks, analyzing also key mathematical needed computation photonic hardware. Then, we investigate, an intuitive analysis, interdependence bit precision efficiency analog circuitry, discussing opportunities PNNs. Followingly, performance overview architectures, weight technologies, activation functions presented, summarizing impact speed, scalability, power consumption. Finally, provide holistic optics-informed framework incorporates physical properties blocks process order improve classification accuracy effectively elevate neuromorphic hardware high-performance DL computational settings.

Language: Английский

---

Orbital angular momentum-mediated machine learning for high-accuracy mode-feature encoding

Light Science & Applications, Journal Year: 2024, Volume and Issue: 13(1)

Published: Feb. 14, 2024

Machine learning with optical neural networks has featured unique advantages of the information processing including high speed, ultrawide bandwidths and low energy consumption because dimensions (time, space, wavelength, polarization) could be utilized to increase degree freedom. However, due lack capability extract features in orbital angular momentum (OAM) domain, theoretically unlimited OAM states have never been exploited represent signal input/output nodes network model. Here, we demonstrate OAM-mediated machine an all-optical convolutional (CNN) based on Laguerre-Gaussian (LG) beam modes diverse diffraction losses. The proposed CNN architecture is composed a trainable mode-dispersion impulse as kernel for feature extraction, deep-learning diffractive layers classifier. resultant selectivity can applied mode-feature encoding, leading accuracy 97.2% MNIST database through detecting weighting coefficients encoded modes, well resistance eavesdropping point-to-point free-space transmission. Moreover, extending target into multiplexed states, realize dimension reduction anomaly detection 85%. Our work provides deep insight mechanism spatial basis, which further improve performances various machine-vision tasks by constructing unsupervised learning-based auto-encoder.

Language: Английский

---

Partial coherence enhances parallelized photonic computing

Nature, Journal Year: 2024, Volume and Issue: 632(8023), P. 55 - 62

Published: July 31, 2024

Abstract Advancements in optical coherence control 1–5 have unlocked many cutting-edge applications, including long-haul communication, light detection and ranging (LiDAR) tomography 6–8 . Prevailing wisdom suggests that using more coherent sources leads to enhanced system performance device functionalities 9–11 Our study introduces a photonic convolutional processing takes advantage of partially boost computing parallelism without substantially sacrificing accuracy, potentially enabling larger-size tensor cores. The reduction the degree optimizes bandwidth use system. This breakthrough challenges traditional belief is essential or even advantageous integrated accelerators, thereby with less rigorous feedback thermal-management requirements for high-throughput computing. Here we demonstrate such two platforms applications: core phase-change-material memories delivers parallel convolution operations classify gaits ten patients Parkinson’s disease 92.2% accuracy (92.7% theoretically) silicon embedded electro-absorption modulators (EAMs) facilitate 0.108 tera per second (TOPS) classifying Modified National Institute Standards Technology (MNIST) handwritten digits dataset 92.4% (95.0% theoretically).

Language: Английский

---

Computational Complexity Optimization of Neural Network-Based Equalizers in Digital Signal Processing: A Comprehensive Approach

Journal of Lightwave Technology, Journal Year: 2024, Volume and Issue: 42(12), P. 4177 - 4201

Published: April 10, 2024

Experimental results based on offline processing reported at optical conferences increasingly rely neural network-based equalizers for accurate data recovery. However, achieving low-complexity implementations that are efficient real-time digital signal remains a challenge. This paper addresses this critical need by proposing systematic approach to designing and evaluating network equalizers. Our focuses three key phases: training, inference, hardware synthesis. We provide comprehensive review of existing methods reducing complexity in each phase, enabling informed choices during design. For the training inference phases, we introduce novel methodology quantifying complexity. includes new metrics bridge software-to-hardware considerations, revealing relationship between specific architectures hyperparameters. guide calculation these both feed-forward recurrent layers, highlighting appropriate choice depending application's focus (software or hardware). Finally, demonstrate practical benefits our approach, showcase how computational can be significantly reduced measured teacher (biLSTM+CNN) student (1D-CNN) different scenarios. work aims standardize estimation optimization networks applied processing, paving way more deployable communication systems.

Language: Английский

---

A guidance to intelligent metamaterials and metamaterials intelligence

Nature Communications, Journal Year: 2025, Volume and Issue: 16(1)

Published: Jan. 29, 2025

Language: Английский

---

Integrated Photonic Neural Networks for Equalizing Optical Communication Signals: A Review

Photonics, Journal Year: 2025, Volume and Issue: 12(1), P. 39 - 39

Published: Jan. 4, 2025

The demand for high-capacity communication systems has grown exponentially in recent decades, constituting a technological field constant change. Data transmission at high rates, reaching tens of Gb/s, and over distances that can reach hundreds kilometers, still faces barriers to improvement, such as distortions the transmitted signals. Such include chromatic dispersion, which causes broadening pulse. Therefore, development solutions adequate recovery signals distorted by complex dynamics channel currently constitutes an open problem since, despite existence well-known efficient equalization techniques, these have limitations terms processing time, hardware complexity, especially energy consumption. In this scenario, paper discusses emergence photonic neural networks promising alternative equalizing optical Thus, review focuses on applications, challenges, opportunities implementing integrated scenario signal equalization. main work carried out, ongoing investigations, possibilities new research directions are also addressed. From review, it be concluded perceptron perform slightly better greater than reservoir computing networks, but with lower data rates. It is important emphasize photonics been growing years, so beyond scope address all existing applications networks.

Language: Английский

---

Progress on intelligent metasurfaces for signal relay, transmitter, and processor

Light Science & Applications, Journal Year: 2025, Volume and Issue: 14(1)

Published: Feb. 25, 2025

Abstract Pursuing higher data rate with limited spectral resources is a longstanding topic that has triggered the fast growth of modern wireless communication techniques. However, massive deployment active nodes to compensate for propagation loss necessitates high hardware expenditure, energy consumption, and maintenance cost, as well complicated network interference issues. Intelligent metasurfaces, composed number subwavelength passive or meta-atoms, have recently found be new paradigm actively reshape environment in green way, distinct from conventional works passively adapt surrounding. In this review, we offer unified perspective on how intelligent metasurfaces can facilitate three manners: signal relay, transmitter, processor. We start by basic modeling channel evolution passive, metasurfaces. Integrated various deep learning algorithms, cater ever-changing environments without human intervention. Then, overview specific experimental advancements using conclude identifying key issues practical implementations surveying directions, such gain knowledge migration.

Language: Английский

---

Photonic (computational) memories: tunable nanophotonics for data storage and computing

Nanophotonics, Journal Year: 2022, Volume and Issue: 11(17), P. 3823 - 3854

Published: May 13, 2022

The exponential growth of information stored in data centers and computational power required for various data-intensive applications, such as deep learning AI, call new strategies to improve or move beyond the traditional von Neumann architecture. Recent achievements storage computation optical domain, enabling energy-efficient, fast, high-bandwidth processing, show great potential photonics overcome bottleneck reduce energy wasted Joule heating. Optically readable memories are fundamental this process, while light-based has traditionally (and commercially) employed free-space optics, recent developments photonic integrated circuits (PICs) nano-materials have opened doors opportunities on-chip. Photonic yet rival their electronic digital counterparts density; however, inherent analog nature ultrahigh bandwidth make them ideal unconventional computing strategies. Here, we review emerging nanophotonic devices that possess memory capabilities by elaborating on tunable mechanisms evaluating terms scalability device performance. Moreover, discuss progress large-scale architectures arrays primarily based

Language: Английский

---