Optical neural networks: progress and challenges

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

Published: Sept. 20, 2024

Language: Английский

---

Multimodal deep learning using on-chip diffractive optics with in situ training capability

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: July 23, 2024

Abstract Multimodal deep learning plays a pivotal role in supporting the processing and of diverse data types within realm artificial intelligence generated content (AIGC). However, most photonic neuromorphic processors for can only handle single modality (either vision or audio) due to lack abundant parameter training optical domain. Here, we propose demonstrate trainable diffractive neural network (TDONN) chip based on on-chip optics with massive tunable elements address these constraints. The TDONN includes one input layer, five hidden layers, output forward propagation is required obtain inference results without frequent optical-electrical conversion. customized stochastic gradient descent algorithm drop-out mechanism are developed neurons realize situ fast convergence achieves potential throughput 217.6 tera-operations per second (TOPS) high computing density (447.7 TOPS/mm 2 ), system-level energy efficiency (7.28 TOPS/W), low latency (30.2 ps). has successfully implemented four-class classification different modalities (vision, audio, touch) achieve 85.7% accuracy multimodal test sets. Our work opens up new avenue integrated processors, providing solution low-power AI large models using technology.

Language: Английский

---

Multimode diffractive optical neural network

Advanced Photonics Nexus, Journal Year: 2024, Volume and Issue: 3(02)

Published: March 8, 2024

On-chip diffractive optical neural networks (DONNs) bring the advantages of parallel processing and low energy consumption. However, an accurate representation field's evolution in structure cannot be provided using previous diffraction-based analysis method. Moreover, loss caused by open boundaries poses challenges to applications. A multimode DONN architecture based on a more precise eigenmode method is proposed. We have constructed universal library input, output, metaline structures utilizing this method, realized composed from library. On designed DONNs with only one layer metaline, classification task Iris plants dataset verified accuracy 90% blind test dataset, performance one-bit binary adder also validated. Compared architectures, exhibits compact design higher efficiency.

Language: Английский

---

Interdisciplinary advances in microcombs: bridging physics and information technology

eLight, Journal Year: 2024, Volume and Issue: 4(1)

Published: Oct. 10, 2024

Abstract The advancement of microcomb sources, which serve as a versatile and powerful platform for various time–frequency measurements, have spurred widespread interest across disciplines. Their uses span coherent optical microwave communications, atomic clocks, high-precision LiDARs, spectrometers, frequency synthesizers. Recent breakthroughs in fabricating micro-cavities, along with the excitation control microcombs, broadened their applications, bridging gap between physical exploration practical engineering systems. These developments pave way pioneering approaches both classical quantum information sciences. In this review article, we conduct thorough examination latest strategies related to enhancement functionalization schemes, cutting-edge applications that cover signal generation, data transmission, analysis, gathering, processing computation. Additionally, provide in-depth evaluations microcomb-based methodologies tailored variety applications. To conclude, consider current state research suggest prospective roadmap could transition technology from laboratory settings broader real-world

Language: Английский

---

Multimodal In‐Sensor Computing System Using Integrated Silicon Photonic Convolutional Processor

Advanced Science, Journal Year: 2024, Volume and Issue: unknown

Published: Oct. 28, 2024

Abstract Photonic integrated circuits offer miniaturized solutions for multimodal spectroscopic sensory systems by leveraging the simultaneous interaction of light with temperature, chemicals, and biomolecules, among others. The data is complex has huge volume high redundancy, thus requiring communication bandwidth associated power consumption to transfer data. To circumvent this cost, photonic sensor processor are brought into intimacy propose a in‐sensor computing system using an silicon convolutional processor. A microring resonator crossbar array used as implement operation 5‐bit accuracy, validated through image edge detection tasks. Further integrating sensor, in situ processing demonstrated, achieving classification protein species different types concentrations at various temperatures. accuracy 97.58% across 45 classes achieved. demonstrates feasibility processors sensors enhance capability devices edge.

Language: Английский

---

Enhancing Simplified Chinese Poetry Comprehension in LLaMA-7B: A Novel Approach to Mimic Mixture of Experts Effect

Research Square (Research Square), Journal Year: 2023, Volume and Issue: unknown

Published: Dec. 20, 2023

Abstract This study explored the potential of manual augmentation in enhancing comprehension and translation capabilities large language models, specifically focusing on LLaMA-7B model context Chinese poetry. poetry, with its rich cultural historical complexities, presents a unique challenge for AI models predominantly trained modern English datasets. Our research introduced novel approach by manually augmenting LLaMA-7B, emulating mixture-of-experts integration. method involved integrating specialized linguistic processing units, which significantly improved model's ability to interpret complex tonal patterns, metaphorical richness, allusions inherent We conducted rigorous evaluations, measuring augmented performance against expert translations noting 23% increase accuracy 37% reduction semantic hallucinations. findings not only demonstrated efficacy bridging gap between demands classical literary texts, but also opened new avenues applying similar techniques other culturally-rich languages. underscored importance contextual awareness processing, marking step towards more advanced culturally sensitive models.

Language: Английский

---

Nonlinear photonics on integrated platforms

Nanophotonics, Journal Year: 2024, Volume and Issue: 13(18), P. 3253 - 3278

Published: June 26, 2024

Nonlinear photonics has unveiled new avenues for applications in metrology, spectroscopy, and optical communications. Recently, there been a surge of interest integrated platforms, attributed to their fundamental benefits, including compatibility with complementary metal-oxide semiconductor (CMOS) processes, reduced power consumption, compactness, cost-effectiveness. This paper provides comprehensive review the key nonlinear effects material properties utilized platforms. It discusses significant achievements supercontinuum generation, phenomenon. Additionally, evolution chip-based frequency combs is reviewed, highlighting recent pivotal works across four main categories. The also examines advances on-chip switching, computing, signal processing, microwave quantum applications. Finally, it perspectives on development challenges offering insights into future directions this rapidly evolving field.

Language: Английский

---

Unlocking High‐Speed and Energy‐Efficiency: Integrated Convolution Processing on Thin‐Film Lithium Niobate

Laser & Photonics Review, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 10, 2025

Abstract Optical neural networks (ONNs) have emerged as high‐performance network accelerators, owing to its broad bandwidth and low power consumption. However, most current ONN architectures still struggle fully leverage their advantages in processing speed energy efficiency. Here, we demonstrate a large‐scale, ultra‐high‐speed, low‐power distributed parallel computing architecture, implemented on thin‐film lithium niobate platform. It can encode image information at modulation rate of 128 Gbaud perform 16 2 × convolution kernel operations, achieving 8.190 trillion multiply‐accumulate operations per second (TMACs/s) with efficiency 4.55 tera watt (Tops/W). This work conducts proof‐of‐concept experiments for edge detection three different ten‐class dataset recognitions, showing performance comparable digital computers. Thanks excellent scalability, high speed, consumption, the integrated optical architecture shows great potential much more sophisticated tasks demanding applications, such autonomous driving video action recognition.

Language: Английский

---

Complex-valued matrix-vector multiplication using a scalable coherent photonic processor

Science Advances, Journal Year: 2025, Volume and Issue: 11(14)

Published: April 4, 2025

Matrix-vector multiplication is a fundamental operation in modern signal processing and artificial intelligence. Developing chip-scale photonic matrix-vector processor (MVMP) offers the potential for notably enhanced computing speed energy efficiency beyond microelectronics. Here, we propose demonstrate 16-channel programmable on-chip coherent capable of performing complex-valued at 1.28 tera-operations per second (TOPS). Low phase error Mach-Zehnder interferometers mesh ultralow-loss broadened waveguide delay lines are firstly combined optical computing, enabling encoding amplitude information, along with high-speed detection. The proposed MVMP demonstrates high flexibility implementing various functions, including arbitrary matrix transformation, parallel image processing, handwritten digital recognition. Our work MVMP’s advantages scalability function flexibility, enabled by low-loss low designs, making substantial advancement large-scale technologies.

Language: Английский

---

Cascaded Micro‐Ring Resonators for Low‐Crosstalk High‐Density Photonic Convolutional Computing

Laser & Photonics Review, Journal Year: 2025, Volume and Issue: unknown

Published: April 26, 2025

Abstract Photonic neural networks (PNNs) based on micro‐ring resonators (MRRs) have attracted significant attention for their compactness and low power consumption. However, there remains substantial room improvement in spectral density network performance. Here, a novel PNN architecture is introduced double‐stage serially coupled ring (DCRRs), incorporating specially designed optoelectronic signal modulation detection circuits, achieving with high density, robustness, accuracy. The DCRR achieves an extinction ratio of 55 dB narrow bandwidth 0.17 nm. Through systematic innovation, it addresses the challenge representing negative numbers caused by non‐negativity light intensity, enabling positive weighting operations using single photodiode‐based architecture. Experimental validation digital cell edge extraction classification tasks demonstrates accuracies above 95%. Compared to single‐ring computing architectures same parameters, this method significantly reduces inter‐channel crosstalk spacing, leading sixfold increase compute 2.48 TOPS/mm 2 . Furthermore, utilizing DCRR‐based nonlinear activation results faster convergence speed higher recognition provides technical foundation high‐density, high‐precision photonic computing.

Language: Английский

---