George Giamougiannis
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
George Giamougiannis has a total of 80 co-authors. Here are the top co-authors by number of shared publications:
- Apostolos Tsakyridis
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
18 shared publications - Miltiadis Moralis-Pegios
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
16 shared publications - Nikos Pleros
Aristotle University of Thessaloniki
16 shared publications - Angelina Totovic
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
8 shared publications - Manos Kirtas
Computational Intelligence, Deep Learning Group, AIIA Lab, Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
7 shared publications - Nikolaos Passalis
Aristotle University of Thessaloniki, Thessaloniki, Greece
7 shared publications
Latest Publications
- 10 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 44 (6) , 2020-2037
Optical Multibeam Beamforming Network Based on a Universal Linear Optical Crossbar Circuit Architecture
This paper presents a theoretical analysis and experimental validation of a recently proposed scalable optical multibeam beamformer architecture based on an Xbar layout, confirming its scalability and robustness to fabrication errors compared to state-of-the-art optical multibeam architectures. It is shown that the Xbar architecture enables efficient routing and combination of optical signals to form distinct, independently controlled radiation patterns. A comprehensive theoretical analysis of the crossbar beamformer underscores its scalability by comparing 4×64 Xbar and Blass matrix architectures using typical phase shifter loss values. The Xbar beamformer exhibits a 23 dB loss, while the Blass matrix reveals a loss exceeding 129 dB which results in a >100 dB gap between the two architectures. Furthermore, the robustness to fabrication errors is analyzed using a beamforming fidelity metric, which in the Xbar architecture exceeds 97% even for a 64×64 array in the presence of significant fabrication imperfections. Experimental validation of the beamformer's capabilities in beam steering and multibeam operation is presented using a 4×4 silicon photonic integrated circuit. Beam scanning is demonstrated by successfully tuning the beamformer's phase shifters to match the transmit position, emulated by proper phasing of the four input signals. Finally, multibeam operation is experimentally demonstrated by successfully distinguishing four distinct signals with BPSK, QPSK, 8PSK, and 16QAM modulation formats arriving simultaneously at the four beamformer input ports, with their phases configured to emulate simultaneous transmission from four orthogonal directions. The results confirm the multibeam beamformer's capability to accurately separate the four signals, effectively resolving them from their respective spatial directions, achieving EVM values of 28.6% for BPSK, 26.5% for QPSK, 25.0% for 8PSK, and 12.9% for 16QAM.
- 10 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 44 (6) , 2020-2037
Optical Multibeam Beamforming Network Based on a Universal Linear Optical Crossbar Circuit Architecture
This paper presents a theoretical analysis and experimental validation of a recently proposed scalable optical multibeam beamformer architecture based on an Xbar layout, confirming its scalability and robustness to fabrication errors compared to state-of-the-art optical multibeam architectures. It is shown that the Xbar architecture enables efficient routing and combination of optical signals to form distinct, independently controlled radiation patterns. A comprehensive theoretical analysis of the crossbar beamformer underscores its scalability by comparing 4×64 Xbar and Blass matrix architectures using typical phase shifter loss values. The Xbar beamformer exhibits a 23 dB loss, while the Blass matrix reveals a loss exceeding 129 dB which results in a >100 dB gap between the two architectures. Furthermore, the robustness to fabrication errors is analyzed using a beamforming fidelity metric, which in the Xbar architecture exceeds 97% even for a 64×64 array in the presence of significant fabrication imperfections. Experimental validation of the beamformer's capabilities in beam steering and multibeam operation is presented using a 4×4 silicon photonic integrated circuit. Beam scanning is demonstrated by successfully tuning the beamformer's phase shifters to match the transmit position, emulated by proper phasing of the four input signals. Finally, multibeam operation is experimentally demonstrated by successfully distinguishing four distinct signals with BPSK, QPSK, 8PSK, and 16QAM modulation formats arriving simultaneously at the four beamformer input ports, with their phases configured to emulate simultaneous transmission from four orthogonal directions. The results confirm the multibeam beamformer's capability to accurately separate the four signals, effectively resolving them from their respective spatial directions, achieving EVM values of 28.6% for BPSK, 26.5% for QPSK, 25.0% for 8PSK, and 12.9% for 16QAM.
- Open Access
- 14 November 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (46) , eadx1657
Digital and plasmonic artificial neural networks—Improved nonlinear signal processing at high speed and low complexity
Transmission at ever higher data rates increasingly demands more advanced digital signal processing techniques, raising both power consumption and operational costs. Here, we introduce a photonic/plasmonic artificial neural network (ANN) using plasmonic modulators to directly mitigate nonlinear signal distortions carried by an optical carrier. This first-of-its-kind plasmonic ANN achieves an ultracompact footprint and high-speed operation and markedly reduces the need for electronic processing. We compare our plasmonic ANN against a traditional digital feed-forward equalizer and a Volterra series, as well as the corresponding digital ANN. The results demonstrate that an astonishingly small ANN outperforms classical equalizers by attaining higher SNR at smaller computational effort. While the digital ANN offers an ideal implementation, executing the ANN on our first plasmonic chip already shows remarkable equalization performance with minimal components. The findings reveal a path toward ultracompact, high-speed, power-efficient, low-latency alternatives to conventional signal processing.
- 14 November 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (46) , eadx1657
Open AccessDigital and plasmonic artificial neural networks—Improved nonlinear signal processing at high speed and low complexity
Transmission at ever higher data rates increasingly demands more advanced digital signal processing techniques, raising both power consumption and operational costs. Here, we introduce a photonic/plasmonic artificial neural network (ANN) using plasmonic modulators to directly mitigate nonlinear signal distortions carried by an optical carrier. This first-of-its-kind plasmonic ANN achieves an ultracompact footprint and high-speed operation and markedly reduces the need for electronic processing. We compare our plasmonic ANN against a traditional digital feed-forward equalizer and a Volterra series, as well as the corresponding digital ANN. The results demonstrate that an astonishingly small ANN outperforms classical equalizers by attaining higher SNR at smaller computational effort. While the digital ANN offers an ideal implementation, executing the ANN on our first plasmonic chip already shows remarkable equalization performance with minimal components. The findings reveal a path toward ultracompact, high-speed, power-efficient, low-latency alternatives to conventional signal processing.
- Open Access
- 9 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Quantum Electronics
- Vol. 31 (3: AI/ML I) , 1-26
Large-Scale Integrated Photonic Device Platform for Energy-Efficient AI/ML Accelerators
The convergence of deep learning and Big Data has spurred significant interest in developing novel hardware that can run large artificial intelligence (AI) workloads more efficiently. Over the last several years, silicon photonics has emerged as a disruptive technology for next-generation accelerators for machine learning (ML). More recently, the heterogeneous integration of III-V compound semiconductors has opened the door to integrating lasers and semiconductor optical amplifiers at wafer-scale, enabling the scaling of the size, density, and complexity of silicon photonic integrated circuits (PICs). Furthermore, using this technology, all of the individual components required to execute the operations within a neural network are available and can be integrated on the same PIC. Here, we review our innovations of an energy-efficient and scalable silicon photonic platform serving as the underlying foundation for next-generation AI accelerator hardware.
- 9 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Quantum Electronics
- Vol. 31 (3: AI/ML I) , 1-26
Open AccessLarge-Scale Integrated Photonic Device Platform for Energy-Efficient AI/ML Accelerators
The convergence of deep learning and Big Data has spurred significant interest in developing novel hardware that can run large artificial intelligence (AI) workloads more efficiently. Over the last several years, silicon photonics has emerged as a disruptive technology for next-generation accelerators for machine learning (ML). More recently, the heterogeneous integration of III-V compound semiconductors has opened the door to integrating lasers and semiconductor optical amplifiers at wafer-scale, enabling the scaling of the size, density, and complexity of silicon photonic integrated circuits (PICs). Furthermore, using this technology, all of the individual components required to execute the operations within a neural network are available and can be integrated on the same PIC. Here, we review our innovations of an energy-efficient and scalable silicon photonic platform serving as the underlying foundation for next-generation AI accelerator hardware.
- Open Access
- 7 October 2024
- Published by Optica Publishing Group in Optics Express
- Vol. 32 (22) , 37968-37983
Ring-assisted, racetrack, and Mach-Zehnder modulator: which one offers the lowest voltages and chirp-free operation?
This study presents a comparison between resonant and non-resonant electro-optical modulator configurations. The focus lies on finding the configuration with the highest modulation amplitude at the lowest drive voltage while achieving a large electro-optical bandwidth. It is found that the ring-assisted Mach-Zehnder modulator (RaMZM) offers chirp-free and resonantly enhanced modulation without bandwidth limitations imposed by the ring. In contrast, a racetrack modulator (RTM) offers resonant enhancement at the cost of a chirped modulated signal and with a bandwidth limitation. The traditional non-resonant Mach-Zehnder modulator (MZM) configuration requires higher modulation voltages while offering chirp-free operation with a flat frequency response. The RaMZM, therefore, looks like an ideal candidate for encoding information in backbone networks where small modulation voltages and perfect control over the phase of a signal are needed. The results are supported by experiments that show driverless plasmonic modulation at 220 GBaud 2PAM, 160 GBaud 4PAM and 100 GBaud 8PAM with record low peak voltages of 0.5 V.
- 7 October 2024
- Published by Optica Publishing Group in Optics Express
- Vol. 32 (22) , 37968-37983
Open AccessRing-assisted, racetrack, and Mach-Zehnder modulator: which one offers the lowest voltages and chirp-free operation?
This study presents a comparison between resonant and non-resonant electro-optical modulator configurations. The focus lies on finding the configuration with the highest modulation amplitude at the lowest drive voltage while achieving a large electro-optical bandwidth. It is found that the ring-assisted Mach-Zehnder modulator (RaMZM) offers chirp-free and resonantly enhanced modulation without bandwidth limitations imposed by the ring. In contrast, a racetrack modulator (RTM) offers resonant enhancement at the cost of a chirped modulated signal and with a bandwidth limitation. The traditional non-resonant Mach-Zehnder modulator (MZM) configuration requires higher modulation voltages while offering chirp-free operation with a flat frequency response. The RaMZM, therefore, looks like an ideal candidate for encoding information in backbone networks where small modulation voltages and perfect control over the phase of a signal are needed. The results are supported by experiments that show driverless plasmonic modulation at 220 GBaud 2PAM, 160 GBaud 4PAM and 100 GBaud 8PAM with record low peak voltages of 0.5 V.
- Top Cited PapersOpen Access
- 27 June 2024
- Published by Springer Nature in Nature Communications
- Vol. 15 (1) , 1-8
Perfect linear optics using silicon photonics
Recently there has been growing interest in using photonics to perform the linear algebra operations of neuromorphic and quantum computing applications, aiming at harnessing silicon photonics’ (SiPho) high-speed and energy-efficiency credentials. Accurately mapping, however, a matrix into optics remains challenging, since state-of-the-art optical architectures are sensitive to fabrication imperfections. This leads to reduced fidelity that degrades as the insertion losses of the optical matrix nodes or the matrix dimensions increase. In this work, we present the experimental deployment of a 4 × 4 coherent crossbar (Xbar) as a silicon chip and validate experimentally its theoretically predicted fidelity restoration credentials. We demonstrate the experimental implementation of 10,000 arbitrary linear transformations achieving a record-high fidelity of 99.997% ± 0.002, limited mainly by the measurement equipment. Our work represents an integrated optical circuit providing almost unity and loss-independent fidelity in the realization of arbitrary matrices, highlighting light’s credentials in resolving complex computations.
- 27 June 2024
- Published by Springer Nature in Nature Communications
- Vol. 15 (1) , 1-8
Top Cited PapersOpen AccessPerfect linear optics using silicon photonics
Recently there has been growing interest in using photonics to perform the linear algebra operations of neuromorphic and quantum computing applications, aiming at harnessing silicon photonics’ (SiPho) high-speed and energy-efficiency credentials. Accurately mapping, however, a matrix into optics remains challenging, since state-of-the-art optical architectures are sensitive to fabrication imperfections. This leads to reduced fidelity that degrades as the insertion losses of the optical matrix nodes or the matrix dimensions increase. In this work, we present the experimental deployment of a 4 × 4 coherent crossbar (Xbar) as a silicon chip and validate experimentally its theoretically predicted fidelity restoration credentials. We demonstrate the experimental implementation of 10,000 arbitrary linear transformations achieving a record-high fidelity of 99.997% ± 0.002, limited mainly by the measurement equipment. Our work represents an integrated optical circuit providing almost unity and loss-independent fidelity in the realization of arbitrary matrices, highlighting light’s credentials in resolving complex computations.
- 17 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 42 (22) , 7825-7833
Scaling Photonic Neural Networks: A Silicon Photonic GeMM Leveraging a Time-space Multiplexed Xbar
We demonstrate a silicon photonic Xbar-based general matrix multiplier (Xbar GeMM) for optical neural network (NN) applications, utilizing a hybrid time-space multiplexing scheme for supporting matrix dimensions far beyond the dimensions of the Xbar circuit. We present the operational principle of the silicon photonic accelerator that is capable of merging space and time division multiplexing techniques through the use of high-speed input and weighting nodes within a coherent M × N Xbar. The proposed scheme was demonstrated experimentally using a 2×2 Xbar that employs electro-absorption modulators (EAM) with 56 GHz bandwidth both at its input signal vector and its weight matrix modulation stages. Its experimental validation as a photonic GeMM engine was performed for 5, 10, 20, 30 and 50 GBd compute rates and was benchmarked as a NN classifier for the IRIS dataset, successfully executing a total number of 2100 products over a 2×2 matrix hardware with an accuracy up to 93.3%. All SiGe EAMs were driven by high-speed electrical signals with a peak-to-peak voltage ranging between 0.9-1.2 V, suggesting a strong potential for a photonic engine that will be capable to perform with CMOS-compatible driving voltages. Finally, we discuss the pros and cons of the proposed hybrid multiplexing scheme, concluding to a thorough system performance and energy efficiency analysis.
- 17 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 42 (22) , 7825-7833
Scaling Photonic Neural Networks: A Silicon Photonic GeMM Leveraging a Time-space Multiplexed Xbar
We demonstrate a silicon photonic Xbar-based general matrix multiplier (Xbar GeMM) for optical neural network (NN) applications, utilizing a hybrid time-space multiplexing scheme for supporting matrix dimensions far beyond the dimensions of the Xbar circuit. We present the operational principle of the silicon photonic accelerator that is capable of merging space and time division multiplexing techniques through the use of high-speed input and weighting nodes within a coherent M × N Xbar. The proposed scheme was demonstrated experimentally using a 2×2 Xbar that employs electro-absorption modulators (EAM) with 56 GHz bandwidth both at its input signal vector and its weight matrix modulation stages. Its experimental validation as a photonic GeMM engine was performed for 5, 10, 20, 30 and 50 GBd compute rates and was benchmarked as a NN classifier for the IRIS dataset, successfully executing a total number of 2100 products over a 2×2 matrix hardware with an accuracy up to 93.3%. All SiGe EAMs were driven by high-speed electrical signals with a peak-to-peak voltage ranging between 0.9-1.2 V, suggesting a strong potential for a photonic engine that will be capable to perform with CMOS-compatible driving voltages. Finally, we discuss the pros and cons of the proposed hybrid multiplexing scheme, concluding to a thorough system performance and energy efficiency analysis.
- 16 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 42 (10) , 3652-3660
Programmable Tanh- and ELU-Based Photonic Neurons in Optics-Informed Neural Networks
We demonstrate an integrated opto-electronic (ΟΕ) device that can be programmed to provide a set of nonlinear activation functions (AFs) and present its operation within programmable tanh- and ELU-based photonic neurons at line rates up to 10 GBd. The OE activation module provides a set of well-known activation functions that are typically used in DL training models, including the tanh-, ELU- and inverted ELU-like functions. Its performance is experimentally evaluated when incorporated in a 4-input wavelength division multiplexed (WDM) photonic neuron and operating with non-deterministic data patterns, providing “noisy” tanh, ELU and inverted ELU AFs with an error-distribution that has in all cases a standard deviation of <0.49. We also evaluate the trainability of these “noisy” AFs and present for the first time an optics-informed training framework that incorporates the pattern-induced AF variations into the training process, yielding the first noise-aware training scheme where the noise emerges at the nonlinear AF NN segment. The performance analysis of the optics-informed training framework for all three AFs was carried out via Deep Learning setups suitable for classifying the Fashion MNIST and the CIFAR-10 datasets. This analysis has shown that the employment of traditional training schemes leads to significant accuracy degradations, which can be, however, almost completely waived when employing the optics-informed training framework, leading to accuracy values that are almost identical to the reference accuracy values obtained when ideal and noise-less AFs are used.
- 16 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 42 (10) , 3652-3660
Programmable Tanh- and ELU-Based Photonic Neurons in Optics-Informed Neural Networks
We demonstrate an integrated opto-electronic (ΟΕ) device that can be programmed to provide a set of nonlinear activation functions (AFs) and present its operation within programmable tanh- and ELU-based photonic neurons at line rates up to 10 GBd. The OE activation module provides a set of well-known activation functions that are typically used in DL training models, including the tanh-, ELU- and inverted ELU-like functions. Its performance is experimentally evaluated when incorporated in a 4-input wavelength division multiplexed (WDM) photonic neuron and operating with non-deterministic data patterns, providing “noisy” tanh, ELU and inverted ELU AFs with an error-distribution that has in all cases a standard deviation of <0.49. We also evaluate the trainability of these “noisy” AFs and present for the first time an optics-informed training framework that incorporates the pattern-induced AF variations into the training process, yielding the first noise-aware training scheme where the noise emerges at the nonlinear AF NN segment. The performance analysis of the optics-informed training framework for all three AFs was carried out via Deep Learning setups suitable for classifying the Fashion MNIST and the CIFAR-10 datasets. This analysis has shown that the employment of traditional training schemes leads to significant accuracy degradations, which can be, however, almost completely waived when employing the optics-informed training framework, leading to accuracy values that are almost identical to the reference accuracy values obtained when ideal and noise-less AFs are used.
- Top Cited PapersOpen Access
- 1 January 2024
- Published by AIP Publishing in APL Photonics
- Vol. 9 (1)
Photonic neural networks and optics-informed deep learning fundamentals
The recent explosive compute growth, mainly fueled by the boost of artificial intelligence (AI) and deep neural networks (DNNs), is currently instigating the demand for a novel computing paradigm that can overcome the insurmountable barriers imposed by conventional electronic computing architectures. Photonic neural networks (PNNs) implemented on silicon integration platforms stand out as a promising candidate to endow neural network (NN) hardware, offering the potential for energy efficient and ultra-fast computations through the utilization of the unique primitives of photonics, i.e., energy efficiency, THz bandwidth, and low-latency. Thus far, several demonstrations have revealed the huge potential of PNNs in performing both linear and non-linear NN operations at unparalleled speed and energy consumption metrics. Transforming this potential into a tangible reality for deep learning (DL) applications requires, however, a deep understanding of the basic PNN principles, requirements, and challenges across all constituent architectural, technological, and training aspects. In this Tutorial, we, initially, review the principles of DNNs along with their fundamental building blocks, analyzing also the key mathematical operations needed for their computation in photonic hardware. Then, we investigate, through an intuitive mathematical analysis, the interdependence of bit precision and energy efficiency in analog photonic circuitry, discussing the opportunities and challenges of PNNs. Followingly, a performance overview of PNN architectures, weight technologies, and activation functions is presented, summarizing their impact in speed, scalability, and power consumption. Finally, we provide a holistic overview of the optics-informed NN training framework that incorporates the physical properties of photonic building blocks into the training process in order to improve the NN classification accuracy and effectively elevate neuromorphic photonic hardware into high-performance DL computational settings.
- 1 January 2024
- Published by AIP Publishing in APL Photonics
- Vol. 9 (1)
Top Cited PapersOpen AccessPhotonic neural networks and optics-informed deep learning fundamentals
The recent explosive compute growth, mainly fueled by the boost of artificial intelligence (AI) and deep neural networks (DNNs), is currently instigating the demand for a novel computing paradigm that can overcome the insurmountable barriers imposed by conventional electronic computing architectures. Photonic neural networks (PNNs) implemented on silicon integration platforms stand out as a promising candidate to endow neural network (NN) hardware, offering the potential for energy efficient and ultra-fast computations through the utilization of the unique primitives of photonics, i.e., energy efficiency, THz bandwidth, and low-latency. Thus far, several demonstrations have revealed the huge potential of PNNs in performing both linear and non-linear NN operations at unparalleled speed and energy consumption metrics. Transforming this potential into a tangible reality for deep learning (DL) applications requires, however, a deep understanding of the basic PNN principles, requirements, and challenges across all constituent architectural, technological, and training aspects. In this Tutorial, we, initially, review the principles of DNNs along with their fundamental building blocks, analyzing also the key mathematical operations needed for their computation in photonic hardware. Then, we investigate, through an intuitive mathematical analysis, the interdependence of bit precision and energy efficiency in analog photonic circuitry, discussing the opportunities and challenges of PNNs. Followingly, a performance overview of PNN architectures, weight technologies, and activation functions is presented, summarizing their impact in speed, scalability, and power consumption. Finally, we provide a holistic overview of the optics-informed NN training framework that incorporates the physical properties of photonic building blocks into the training process in order to improve the NN classification accuracy and effectively elevate neuromorphic photonic hardware into high-performance DL computational settings.
- 17 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Quantum Electronics
- Vol. 29 (6: Photoni) , 1-10
Programmable Tanh-, ELU-, Sigmoid-, and Sin-Based Nonlinear Activation Functions for Neuromorphic Photonics
We demonstrate a programmable analog opto-electronic (OE) circuit that can be configured to provide a range of nonlinear activation functions for incoherent neuromorphic photonic circuits at up to 10 Gbaud line-rates. We present a set of well-known activation functions that are typically used to train DL models including tanh-, sigmoid-, ReLU- and inverted ReLU-like activations, introducing also a series of novel photonic nonlinear functions that are referred to as Rectified Sine Squared (ReSin), Sine Squared with Exponential tail (ExpSin) and Double Sine Squared. Experimental validation for all these activation functions is performed at 10 Gbaud operation. The ability of the mathematically modelled photonic activation functions to train Deep Neural Networks (DNNs) has been verified through their employment in Deep Learning (DL) models for MNIST and CIFAR10 classification purposes, comparing their performance against corresponding NNs that utilize an ideal ReLU activation function.
- 17 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Quantum Electronics
- Vol. 29 (6: Photoni) , 1-10
Programmable Tanh-, ELU-, Sigmoid-, and Sin-Based Nonlinear Activation Functions for Neuromorphic Photonics
We demonstrate a programmable analog opto-electronic (OE) circuit that can be configured to provide a range of nonlinear activation functions for incoherent neuromorphic photonic circuits at up to 10 Gbaud line-rates. We present a set of well-known activation functions that are typically used to train DL models including tanh-, sigmoid-, ReLU- and inverted ReLU-like activations, introducing also a series of novel photonic nonlinear functions that are referred to as Rectified Sine Squared (ReSin), Sine Squared with Exponential tail (ExpSin) and Double Sine Squared. Experimental validation for all these activation functions is performed at 10 Gbaud operation. The ability of the mathematically modelled photonic activation functions to train Deep Neural Networks (DNNs) has been verified through their employment in Deep Learning (DL) models for MNIST and CIFAR10 classification purposes, comparing their performance against corresponding NNs that utilize an ideal ReLU activation function.
- Open Access
- 9 January 2023
- Published by Wiley in Nanophotonics
- Vol. 12 (5) , 963-973
Analog nanophotonic computing going practical: silicon photonic deep learning engines for tiled optical matrix multiplication with dynamic precision
Analog photonic computing comprises a promising candidate for accelerating the linear operations of deep neural networks (DNNs), since it provides ultrahigh bandwidth, low footprint and low power consumption computing capabilities. However, the confined photonic hardware size, along with the limited bit precision of high‐speed electro‐optical components, impose stringent requirements towards surpassing the performance levels of current digital processors. Herein, we propose and experimentally demonstrate a speed‐optimized dynamic precision neural network (NN) inference via tiled matrix multiplication (TMM) on a low‐radix silicon photonic processor. We introduce a theoretical model that relates the noise figure of a photonic neuron with the bit precision requirements per neural layer. The inference evaluation of an NN trained for the classification of the IRIS dataset is, then, experimentally performed over a silicon coherent photonic neuron that can support optical TMM up to 50 GHz, allowing, simultaneously, for dynamic‐precision calculations. Targeting on a high‐accuracy and speed‐optimized classification performance, we experimentally applied the model‐extracted mixed‐precision NN inference scheme via the respective alteration of the operational compute rates per neural layer. This dynamic‐precision NN inference revealed a 55% decrease in the execution time of the linear operations compared to a fixed‐precision scheme, without degrading its accuracy.
- 9 January 2023
- Published by Wiley in Nanophotonics
- Vol. 12 (5) , 963-973
Open AccessAnalog nanophotonic computing going practical: silicon photonic deep learning engines for tiled optical matrix multiplication with dynamic precision
Analog photonic computing comprises a promising candidate for accelerating the linear operations of deep neural networks (DNNs), since it provides ultrahigh bandwidth, low footprint and low power consumption computing capabilities. However, the confined photonic hardware size, along with the limited bit precision of high‐speed electro‐optical components, impose stringent requirements towards surpassing the performance levels of current digital processors. Herein, we propose and experimentally demonstrate a speed‐optimized dynamic precision neural network (NN) inference via tiled matrix multiplication (TMM) on a low‐radix silicon photonic processor. We introduce a theoretical model that relates the noise figure of a photonic neuron with the bit precision requirements per neural layer. The inference evaluation of an NN trained for the classification of the IRIS dataset is, then, experimentally performed over a silicon coherent photonic neuron that can support optical TMM up to 50 GHz, allowing, simultaneously, for dynamic‐precision calculations. Targeting on a high‐accuracy and speed‐optimized classification performance, we experimentally applied the model‐extracted mixed‐precision NN inference scheme via the respective alteration of the operational compute rates per neural layer. This dynamic‐precision NN inference revealed a 55% decrease in the execution time of the linear operations compared to a fixed‐precision scheme, without degrading its accuracy.