Manos Kirtas
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Manos Kirtas has a total of 65 co-authors. Here are the top co-authors by number of shared publications:
- Nikolaos Passalis
Aristotle University of Thessaloniki, Thessaloniki, Greece
25 shared publications - Anastasios Tefas
Aristotle University of Thessaloniki
23 shared publications - Nikos Pleros
Aristotle University of Thessaloniki
17 shared publications - Miltiadis Moralis-Pegios
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
13 shared publications - Apostolos Tsakyridis
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
11 shared publications - George Giamougiannis
School of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
7 shared publications
Latest Publications
- 3 August 2026
- Published by Springer Nature
- p. 585-592
Optics-Informed Long Short Term Memory Cells
- 11 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 10 (3) , 2471-2481
Non-Negative Gated Recurrent Neural Networks
Neural networks have become increasingly popular in recent years due to their ability to efficiently solve a wide range of complex problems, including computer vision, machine translation, and speech recognition. The core components of neural networks are their learnable parameters, which consist of their weights and biases. In typical neural networks, these parameters can be positive or negative, allowing models to learn complicated patterns in data. However, having both positive and negative values in a network introduces challenges in cases of analog hardware deployments. For instance, neuromorphic photonic hardware is energy-efficient but often lacks the ability to represent negative quantities natively. In this work, we propose a transformation method oriented to gated recurrent neural networks, such as Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) models, converting them directly into deployable models, and targeting hardware accelerators that can only represent non-negative quantities. The proposed approach leverages pre-trained networks to efficiently generate equivalent non-negative models, eliminating the need for training non-negative models from scratch, a process that often leads to models with poor performance due to optimization challenges. The proposed method enables one to efficiently deploy gated RNNs to photonic hardware, while also offering additional understanding of model behavior, as highlighted in this work.
- 11 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 10 (3) , 2471-2481
Non-Negative Gated Recurrent Neural Networks
Neural networks have become increasingly popular in recent years due to their ability to efficiently solve a wide range of complex problems, including computer vision, machine translation, and speech recognition. The core components of neural networks are their learnable parameters, which consist of their weights and biases. In typical neural networks, these parameters can be positive or negative, allowing models to learn complicated patterns in data. However, having both positive and negative values in a network introduces challenges in cases of analog hardware deployments. For instance, neuromorphic photonic hardware is energy-efficient but often lacks the ability to represent negative quantities natively. In this work, we propose a transformation method oriented to gated recurrent neural networks, such as Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) models, converting them directly into deployable models, and targeting hardware accelerators that can only represent non-negative quantities. The proposed approach leverages pre-trained networks to efficiently generate equivalent non-negative models, eliminating the need for training non-negative models from scratch, a process that often leads to models with poor performance due to optimization challenges. The proposed method enables one to efficiently deploy gated RNNs to photonic hardware, while also offering additional understanding of model behavior, as highlighted in this work.
- Open Access
- 20 December 2025
- Published by Elsevier in Applied Soft Computing
- Vol. 189
Neuromorphic photonic neural networks for on-chip gas spectra classification
- 20 December 2025
- Published by Elsevier in Applied Soft Computing
- Vol. 189
Open AccessNeuromorphic photonic neural networks for on-chip gas spectra classification
- Open Access
- 24 November 2025
- Published by Elsevier in Neurocomputing
- Vol. 665
Non-negative isomorphic neural networks for efficient accelerators
- 24 November 2025
- Published by Elsevier in Neurocomputing
- Vol. 665
Open AccessNon-negative isomorphic neural networks for efficient accelerators
- Open Access
- 1 November 2025
- Published by AIP Publishing in APL Photonics
- Vol. 10 (11)
A 262 TOPS hyperdimensional photonic AI accelerator powered by a Si3N4 microcomb laser
The ever-increasing volume of data demarcating from the exponential scale of Artificial Intelligence (AI) and Deep Learning (DL) models motivated research into specialized AI accelerators in order to complement digital processors. Photonic Neural Networks (PNNs), with their unique ability to capitalize on the interplay of multiple physical dimensions, including time, wavelength, and space, have been brought forward with a credible promise for boosting computational power and energy efficiency in AI processors. In this article, we experimentally demonstrate a novel multidimensional arrayed waveguide grating router (AWGR)-based photonic AI accelerator that can offload bandwidth-bounded linear algebra while leaving memory hierarchy, control, and nonlinearities to electronics and can execute tensor multiplications at a record-high total computational power of 262 TOPS, offering a ∼24× improvement over the existing waveguide-based optical accelerators. It consists of a 16 × 16 AWGR that exploits the time-, wavelength-, and space-division multiplexing (T-W-SDM) for weight and input encoding, together with an integrated Si3N4-based frequency comb for multi-wavelength generation. The photonic AI accelerator has been experimentally validated in both Fully Connected (FC) and Convolutional NN (CNN) models, with the FC and CNN being trained for DDoS attack identification and MNIST classification, respectively. The experimental inference at 32 Gbaud achieved a Cohen’s kappa score of 0.8652 for DDoS detection and an accuracy of 92.14% for MNIST classification, respectively, closely matching the software performance.
- 1 November 2025
- Published by AIP Publishing in APL Photonics
- Vol. 10 (11)
Open AccessA 262 TOPS hyperdimensional photonic AI accelerator powered by a Si3N4 microcomb laser
The ever-increasing volume of data demarcating from the exponential scale of Artificial Intelligence (AI) and Deep Learning (DL) models motivated research into specialized AI accelerators in order to complement digital processors. Photonic Neural Networks (PNNs), with their unique ability to capitalize on the interplay of multiple physical dimensions, including time, wavelength, and space, have been brought forward with a credible promise for boosting computational power and energy efficiency in AI processors. In this article, we experimentally demonstrate a novel multidimensional arrayed waveguide grating router (AWGR)-based photonic AI accelerator that can offload bandwidth-bounded linear algebra while leaving memory hierarchy, control, and nonlinearities to electronics and can execute tensor multiplications at a record-high total computational power of 262 TOPS, offering a ∼24× improvement over the existing waveguide-based optical accelerators. It consists of a 16 × 16 AWGR that exploits the time-, wavelength-, and space-division multiplexing (T-W-SDM) for weight and input encoding, together with an integrated Si3N4-based frequency comb for multi-wavelength generation. The photonic AI accelerator has been experimentally validated in both Fully Connected (FC) and Convolutional NN (CNN) models, with the FC and CNN being trained for DDoS attack identification and MNIST classification, respectively. The experimental inference at 32 Gbaud achieved a Cohen’s kappa score of 0.8652 for DDoS detection and an accuracy of 92.14% for MNIST classification, respectively, closely matching the software performance.
- 21 August 2025
- Published by Optica Publishing Group in Optics Express
- Vol. 33 (17) , 36960-36972
Time-space-wavelength multiplexed photonic tensor core using WDM SiGe EAM array chiplets
Photonic neural networks (PNNs) are projected to be the next-generation AI platform, driving significant advances in compute energy and area efficiency by leveraging light’s parallelism across space, wavelength, and time. Here we present a photonic integrated tensor core (PITC) architecture that combines an arrayed waveguide grating router (AWGR)-enabled PNN with a SiGe electro-absorption modulator (EAM) array following a chiplet-based silicon photonic (SiPho) accelerator setup for multi-wavelength data and weight encoding. The 8 × 8 PITC architectural layout was experimentally validated in two AI benchmark workloads at 20 Gbaud, successfully classifying the IRIS and MNIST datasets yielding Cohen’s kappa-scores of 0.8438 and 0.7421, respectively.
- 21 August 2025
- Published by Optica Publishing Group in Optics Express
- Vol. 33 (17) , 36960-36972
Time-space-wavelength multiplexed photonic tensor core using WDM SiGe EAM array chiplets
Photonic neural networks (PNNs) are projected to be the next-generation AI platform, driving significant advances in compute energy and area efficiency by leveraging light’s parallelism across space, wavelength, and time. Here we present a photonic integrated tensor core (PITC) architecture that combines an arrayed waveguide grating router (AWGR)-enabled PNN with a SiGe electro-absorption modulator (EAM) array following a chiplet-based silicon photonic (SiPho) accelerator setup for multi-wavelength data and weight encoding. The 8 × 8 PITC architectural layout was experimentally validated in two AI benchmark workloads at 20 Gbaud, successfully classifying the IRIS and MNIST datasets yielding Cohen’s kappa-scores of 0.8438 and 0.7421, respectively.
- 25 April 2025
- Published by Elsevier in Neural Networks
- Vol. 188, 107492
Sign potential-driven multiplicative optimization for robust deep reinforcement learning
- 25 April 2025
- Published by Elsevier in Neural Networks
- Vol. 188, 107492
Sign potential-driven multiplicative optimization for robust deep reinforcement learning
- 21 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 43 (4) , 1773-1785
Reaching the Peta-Computing: 163.8 TOPS Through Multidimensional AWGR-Based Accelerators
The prowess of Artificial Intelligence (AI) and Large Language Models (LLMs) to compute the exponentially growing data, coupled with the digital electronic AI computing systems reaching their physical plateaus, have stimulated extensive research into next-gen AI accelerators. Optical Neural Networks (ONNs) have been among the most favored candidates, since the unparalleled speed of light can provide massive compute parallelism by the physical dimensions of time, wavelength, and space, offering higher computational speeds and lower power consumption. In this paper, we experimentally demonstrate novel time-space-wavelength multiplexed arrayed waveguided gratings router (AWGR)-based architectures exploited as photonic matrix- and tensor-multiplier engines achieving a total computational power of up to 163.8 TOPS. This marks a substantial ∼15x improvement over state-of-the-art waveguide-based optical accelerators. The two demonstrators were tested for three different NN applications, under two experimental testbeds. The matrix-multiplier was evaluated as a fully connected NN (FCNN), classifying images from the FMNIST dataset. An 8×8 AWGR was employed for providing the passive routing, with modulators driven at 10Gbaud, exhibiting a hardware-based inference accuracy of 87.1% with respect to the software. The tensor-multiplier, using a 16x16 AWGR, was evaluated as a FCNN and a convolutional NN (CNN), detecting DDoS attacks and classifying handwritten digits at 20 Gbaud. The experimental results exhibited a Cohen's kappa score of 0.799 for the DDoS detection and a classification accuracy of 93.35%, respectively.
- 21 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 43 (4) , 1773-1785
Reaching the Peta-Computing: 163.8 TOPS Through Multidimensional AWGR-Based Accelerators
The prowess of Artificial Intelligence (AI) and Large Language Models (LLMs) to compute the exponentially growing data, coupled with the digital electronic AI computing systems reaching their physical plateaus, have stimulated extensive research into next-gen AI accelerators. Optical Neural Networks (ONNs) have been among the most favored candidates, since the unparalleled speed of light can provide massive compute parallelism by the physical dimensions of time, wavelength, and space, offering higher computational speeds and lower power consumption. In this paper, we experimentally demonstrate novel time-space-wavelength multiplexed arrayed waveguided gratings router (AWGR)-based architectures exploited as photonic matrix- and tensor-multiplier engines achieving a total computational power of up to 163.8 TOPS. This marks a substantial ∼15x improvement over state-of-the-art waveguide-based optical accelerators. The two demonstrators were tested for three different NN applications, under two experimental testbeds. The matrix-multiplier was evaluated as a fully connected NN (FCNN), classifying images from the FMNIST dataset. An 8×8 AWGR was employed for providing the passive routing, with modulators driven at 10Gbaud, exhibiting a hardware-based inference accuracy of 87.1% with respect to the software. The tensor-multiplier, using a 16x16 AWGR, was evaluated as a FCNN and a convolutional NN (CNN), detecting DDoS attacks and classifying handwritten digits at 20 Gbaud. The experimental results exhibited a Cohen's kappa score of 0.799 for the DDoS detection and a classification accuracy of 93.35%, respectively.
- 1 January 2025
- Published by Springer Nature
- p. 420-432
Using Part-Based Representations for Explainable Deep Reinforcement Learning
- 1 January 2025
- Published by Springer Nature
- p. 420-432
Using Part-Based Representations for Explainable Deep Reinforcement Learning
- 2 December 2024
- Published by Springer Nature
- p. 322-336
Multiplicative RMSprop Using Gradient Normalization for Learning Acceleration
- 2 December 2024
- Published by Springer Nature
- p. 322-336
Multiplicative RMSprop Using Gradient Normalization for Learning Acceleration