Ilker Oguz
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ilker Oguz has a total of 48 co-authors. Here are the top co-authors by number of shared publications:
- Christophe Moser
Laboratory of Applied Photonics Devices, School of Engineering, École Polytechnique Fédérale de Lausanne, CH-1015, Lausanne, Switzerland
7 shared publications Mustafa Yildirim
EPFL Institute of Electrical and Micro Engineering 1015 Lausanne Switzerland
7 shared publications- Demetri Psaltis
Optics Laboratory (LO), École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
6 shared publications Demetri Psaltis
5 shared publications- Niyazi Ulas Dinc
EPFL Institute of Electrical and Micro Engineering 1015 Lausanne Switzerland
5 shared publications Jih-Liang Hsieh
EPFL, Institute of Electrical and Micro Engineering, Lausanne, Switzerland
4 shared publications
Latest Publications
- Open Access
Exploring Expert Specialization through Unsupervised Training in Sparse Mixture of Experts
Understanding the internal organization of neural networks remains a fundamental challenge in deep learning interpretability. We address this challenge by exploring a novel Sparse Mixture of Experts Variational Autoencoder (SMoE-VAE) architecture. We test our model on the QuickDraw dataset, comparing unsupervised expert routing against a supervised baseline guided by ground-truth labels. Surprisingly, we find that unsupervised routing consistently achieves superior reconstruction performance. The experts learn to identify meaningful sub-categorical structures that often transcend human-defined class boundaries. Through t-SNE visualizations and reconstruction analysis, we investigate how MoE models uncover fundamental data structures that are more aligned with the model's objective than predefined labels. Furthermore, our study on the impact of dataset size provides insights into the trade-offs between data quantity and expert specialization, offering guidance for designing efficient MoE architectures.
Open AccessExploring Expert Specialization through Unsupervised Training in Sparse Mixture of Experts
Understanding the internal organization of neural networks remains a fundamental challenge in deep learning interpretability. We address this challenge by exploring a novel Sparse Mixture of Experts Variational Autoencoder (SMoE-VAE) architecture. We test our model on the QuickDraw dataset, comparing unsupervised expert routing against a supervised baseline guided by ground-truth labels. Surprisingly, we find that unsupervised routing consistently achieves superior reconstruction performance. The experts learn to identify meaningful sub-categorical structures that often transcend human-defined class boundaries. Through t-SNE visualizations and reconstruction analysis, we investigate how MoE models uncover fundamental data structures that are more aligned with the model's objective than predefined labels. Furthermore, our study on the impact of dataset size provides insights into the trade-offs between data quantity and expert specialization, offering guidance for designing efficient MoE architectures.
- Top Cited Papers
- 3 September 2025
- Published by Springer Nature in Nature
- Vol. 645 (8079) , 53-61
Training of physical neural networks
Physical neural networks (PNNs) are a class of neural-like networks that make use of analogue physical systems to perform computations. Although at present confined to small-scale laboratory demonstrations, PNNs could one day transform how artificial intelligence (AI) calculations are performed. Could we train AI models many orders of magnitude larger than present ones? Could we perform model inference locally and privately on edge devices? Research over the past few years has shown that the answer to these questions is probably “yes, with enough research”. Because PNNs can make use of analogue physical computations more directly, flexibly and opportunistically than traditional computing hardware, they could change what is possible and practical for AI systems. To do this, however, will require notable progress, rethinking both how AI models work and how they are trained—primarily by considering the problems through the constraints of the underlying hardware physics. To train PNNs, backpropagation-based and backpropagation-free approaches are now being explored. These methods have various trade-offs and, so far, no method has been shown to scale to large models with the same performance as the backpropagation algorithm widely used in deep learning today. However, this challenge has been rapidly changing and a diverse ecosystem of training techniques provides clues for how PNNs may one day be used to create both more efficient and larger-scale realizations of present-scale AI models.
- 3 September 2025
- Published by Springer Nature in Nature
- Vol. 645 (8079) , 53-61
Top Cited PapersTraining of physical neural networks
Physical neural networks (PNNs) are a class of neural-like networks that make use of analogue physical systems to perform computations. Although at present confined to small-scale laboratory demonstrations, PNNs could one day transform how artificial intelligence (AI) calculations are performed. Could we train AI models many orders of magnitude larger than present ones? Could we perform model inference locally and privately on edge devices? Research over the past few years has shown that the answer to these questions is probably “yes, with enough research”. Because PNNs can make use of analogue physical computations more directly, flexibly and opportunistically than traditional computing hardware, they could change what is possible and practical for AI systems. To do this, however, will require notable progress, rethinking both how AI models work and how they are trained—primarily by considering the problems through the constraints of the underlying hardware physics. To train PNNs, backpropagation-based and backpropagation-free approaches are now being explored. These methods have various trade-offs and, so far, no method has been shown to scale to large models with the same performance as the backpropagation algorithm widely used in deep learning today. However, this challenge has been rapidly changing and a diverse ecosystem of training techniques provides clues for how PNNs may one day be used to create both more efficient and larger-scale realizations of present-scale AI models.
- Open Access
- 2 April 2025
- Published by Wiley in Nanophotonics
- Vol. 14 (16) , 2787-2797
Training hybrid neural networks with multimode optical nonlinearities using digital twins
The ability to train ever‐larger neural networks brings artificial intelligence to the forefront of scientific and technical discoveries. However, their exponentially increasing size creates a proportionally greater demand for energy and computational hardware. Incorporating complex physical events in networks as fixed, efficient computation modules can address this demand by decreasing the complexity of trainable layers. Here, we utilize ultrashort pulse propagation in multimode fibers, which perform large‐scale nonlinear transformations, for this purpose. Training the hybrid architecture is achieved through a neural model that differentiably approximates the optical system. The training algorithm updates the neural simulator and backpropagates the error signal over this proxy to optimize layers preceding the optical one. Our experimental results achieve state‐of‐the‐art image classification accuracies and simulation fidelity. Moreover, the framework demonstrates exceptional resistance to experimental drifts. By integrating low‐energy physical systems into neural networks, this approach enables scalable, energy‐efficient AI models with significantly reduced computational demands.
- 2 April 2025
- Published by Wiley in Nanophotonics
- Vol. 14 (16) , 2787-2797
Open AccessTraining hybrid neural networks with multimode optical nonlinearities using digital twins
The ability to train ever‐larger neural networks brings artificial intelligence to the forefront of scientific and technical discoveries. However, their exponentially increasing size creates a proportionally greater demand for energy and computational hardware. Incorporating complex physical events in networks as fixed, efficient computation modules can address this demand by decreasing the complexity of trainable layers. Here, we utilize ultrashort pulse propagation in multimode fibers, which perform large‐scale nonlinear transformations, for this purpose. Training the hybrid architecture is achieved through a neural model that differentiably approximates the optical system. The training algorithm updates the neural simulator and backpropagates the error signal over this proxy to optimize layers preceding the optical one. Our experimental results achieve state‐of‐the‐art image classification accuracies and simulation fidelity. Moreover, the framework demonstrates exceptional resistance to experimental drifts. By integrating low‐energy physical systems into neural networks, this approach enables scalable, energy‐efficient AI models with significantly reduced computational demands.
- Open Access
- 4 January 2025
- Published by Springer Nature in Light: Science & Applications
- Vol. 14 (1) , 1-4
Resource-efficient photonic networks for next-generation AI computing
Current trends in artificial intelligence toward larger models demand a rethinking of both hardware and algorithms. Photonics-based systems offer high-speed, energy-efficient computing units, provided algorithms are designed to exploit photonics’ unique strengths. The recent implementation of cellular automata in photonics demonstrates how a few local interactions can achieve high throughput and precision.
- 4 January 2025
- Published by Springer Nature in Light: Science & Applications
- Vol. 14 (1) , 1-4
Open AccessResource-efficient photonic networks for next-generation AI computing
Current trends in artificial intelligence toward larger models demand a rethinking of both hardware and algorithms. Photonics-based systems offer high-speed, energy-efficient computing units, provided algorithms are designed to exploit photonics’ unique strengths. The recent implementation of cellular automata in photonics demonstrates how a few local interactions can achieve high throughput and precision.
- 20 September 2024
- Published by Optica Publishing Group in Optical Materials Express
- Vol. 14 (10) , 2413-2418
On the implementation of nonlinearities in optical neural networks: opinion
Optical neural networks are expected to be hybrid systems for the foreseeable future combining optics and electronics. The linear portion of the system, consisting of the weighted interconnections between the neurons, can be realized optically. The nonlinear activation functions and the training of the network are implemented with nonlinear optics or electronically.
- 20 September 2024
- Published by Optica Publishing Group in Optical Materials Express
- Vol. 14 (10) , 2413-2418
On the implementation of nonlinearities in optical neural networks: opinion
Optical neural networks are expected to be hybrid systems for the foreseeable future combining optics and electronics. The linear portion of the system, consisting of the weighted interconnections between the neurons, can be realized optically. The nonlinear activation functions and the training of the network are implemented with nonlinear optics or electronically.
- Top Cited PapersOpen Access
- 31 July 2024
- Published by Springer Nature in Nature Photonics
- Vol. 18 (10) , 1076-1082
Nonlinear processing with linear optics
Deep neural networks have achieved remarkable breakthroughs by leveraging multiple layers of data processing to extract hidden representations, albeit at the cost of large electronic computing power. To enhance energy efficiency and speed, the optical implementation of neural networks aims to harness the advantages of optical bandwidth and the energy efficiency of optical interconnections. In the absence of low-power optical nonlinearities, the challenge in the implementation of multilayer optical networks lies in realizing multiple optical layers without resorting to electronic components. Here we present a novel framework that uses multiple scattering, and which is capable of synthesizing programmable linear and nonlinear transformations concurrently at low optical power by leveraging the nonlinear relationship between the scattering potential, represented by data, and the scattered field. Theoretical and experimental investigations show that repeating the data by multiple scattering enables nonlinear optical computing with low-power continuous-wave light. Moreover, we empirically find that scaling of this optical framework follows a power law.
- 31 July 2024
- Published by Springer Nature in Nature Photonics
- Vol. 18 (10) , 1076-1082
Top Cited PapersOpen AccessNonlinear processing with linear optics
Deep neural networks have achieved remarkable breakthroughs by leveraging multiple layers of data processing to extract hidden representations, albeit at the cost of large electronic computing power. To enhance energy efficiency and speed, the optical implementation of neural networks aims to harness the advantages of optical bandwidth and the energy efficiency of optical interconnections. In the absence of low-power optical nonlinearities, the challenge in the implementation of multilayer optical networks lies in realizing multiple optical layers without resorting to electronic components. Here we present a novel framework that uses multiple scattering, and which is capable of synthesizing programmable linear and nonlinear transformations concurrently at low optical power by leveraging the nonlinear relationship between the scattering potential, represented by data, and the scattered field. Theoretical and experimental investigations show that repeating the data by multiple scattering enables nonlinear optical computing with low-power continuous-wave light. Moreover, we empirically find that scaling of this optical framework follows a power law.
- Top Cited PapersOpen Access
- 1 January 2024
- Published by SPIE-Intl Soc Optical Eng in Advanced Photonics
- Vol. 6 (01) , 016002-11
Programming nonlinear propagation for efficient optical learning machines
Machine-learning architectures have come to be dominated by artificial neural networks (ANNs). There are several reasons why this architecture is used so broadly. Initially, their similarity with biological neural networks1 provided strong motivation to explore ANNs. At the same time, the fact that ANNs are universal machines2 that are able to approximate any function breeds confidence that ANNs can carry out useful and difficult tasks. Perhaps most significantly, the fact that error backpropagation3 has proven very effective in training such networks catapulted their application to a wide variety of problems. Ever-larger networks4 have been adopted for tackling challenging tasks.5 Empirically, it has been found that larger networks tend to perform better given a sufficiently large database of training examples. This has led to a “bigger is better” mentality.6 However, the disadvantage of this mentality is the energy required to train and use very large networks. For instance, only training the language model GPT-3, which has 175 billion parameters, consumed 1.3 GWh of electricity, which is the energy required to fully charge 13,000 Tesla Model S cars.7 Optics can help overcome this downside, since light propagation through a nonabsorbing, nonscattering medium is a lossless linear operation.
- 1 January 2024
- Published by SPIE-Intl Soc Optical Eng in Advanced Photonics
- Vol. 6 (01) , 016002-11
Top Cited PapersOpen AccessProgramming nonlinear propagation for efficient optical learning machines
Machine-learning architectures have come to be dominated by artificial neural networks (ANNs). There are several reasons why this architecture is used so broadly. Initially, their similarity with biological neural networks1 provided strong motivation to explore ANNs. At the same time, the fact that ANNs are universal machines2 that are able to approximate any function breeds confidence that ANNs can carry out useful and difficult tasks. Perhaps most significantly, the fact that error backpropagation3 has proven very effective in training such networks catapulted their application to a wide variety of problems. Ever-larger networks4 have been adopted for tackling challenging tasks.5 Empirically, it has been found that larger networks tend to perform better given a sufficiently large database of training examples. This has led to a “bigger is better” mentality.6 However, the disadvantage of this mentality is the energy required to train and use very large networks. For instance, only training the language model GPT-3, which has 175 billion parameters, consumed 1.3 GWh of electricity, which is the energy required to fully charge 13,000 Tesla Model S cars.7 Optics can help overcome this downside, since light propagation through a nonabsorbing, nonscattering medium is a lossless linear operation.
- 3 October 2023
- Published by Optica Publishing Group in Optics Letters
- Vol. 48 (20) , 5249-5252
Forward–forward training of an optical neural network
Neural networks (NNs) have demonstrated remarkable capabilities in various tasks, but their computation-intensive nature demands faster and more energy-efficient hardware implementations. Optics-based platforms, using technologies such as silicon photonics and spatial light modulators, offer promising avenues for achieving this goal. However, training multiple programmable layers together with these physical systems poses challenges, as they are difficult to fully characterize and describe with differentiable functions, hindering the use of error backpropagation algorithm. The recently introduced forward–forward algorithm (FFA) eliminates the need for perfect characterization of the physical learning system and shows promise for efficient training with large numbers of programmable parameters. The FFA does not require backpropagating an error signal to update the weights, rather the weights are updated by only sending information in one direction. The local loss function for each set of trainable weights enables low-power analog hardware implementations without resorting to metaheuristic algorithms or reinforcement learning. In this paper, we present an experiment utilizing multimode nonlinear wave propagation in an optical fiber demonstrating the feasibility of the FFA approach using an optical system. The results show that incorporating optical transforms in multilayer NN architectures trained with the FFA can lead to performance improvements, even with a relatively small number of trainable weights. The proposed method offers a new path to the challenge of training optical NNs and provides insights into leveraging physical transformations for enhancing the NN performance.
- 3 October 2023
- Published by Optica Publishing Group in Optics Letters
- Vol. 48 (20) , 5249-5252
Forward–forward training of an optical neural network
Neural networks (NNs) have demonstrated remarkable capabilities in various tasks, but their computation-intensive nature demands faster and more energy-efficient hardware implementations. Optics-based platforms, using technologies such as silicon photonics and spatial light modulators, offer promising avenues for achieving this goal. However, training multiple programmable layers together with these physical systems poses challenges, as they are difficult to fully characterize and describe with differentiable functions, hindering the use of error backpropagation algorithm. The recently introduced forward–forward algorithm (FFA) eliminates the need for perfect characterization of the physical learning system and shows promise for efficient training with large numbers of programmable parameters. The FFA does not require backpropagating an error signal to update the weights, rather the weights are updated by only sending information in one direction. The local loss function for each set of trainable weights enables low-power analog hardware implementations without resorting to metaheuristic algorithms or reinforcement learning. In this paper, we present an experiment utilizing multimode nonlinear wave propagation in an optical fiber demonstrating the feasibility of the FFA approach using an optical system. The results show that incorporating optical transforms in multilayer NN architectures trained with the FFA can lead to performance improvements, even with a relatively small number of trainable weights. The proposed method offers a new path to the challenge of training optical NNs and provides insights into leveraging physical transformations for enhancing the NN performance.
- Open Access
- 1 October 2023
- Published by AIP Publishing in APL Photonics
- Vol. 8 (10)
Nonlinear optical feature generator for machine learning
Modern machine learning models use an ever-increasing number of parameters to train (175 × 109 parameters for GPT-3) with large datasets to achieve better performance. Optical computing has been rediscovered as a potential solution for large-scale data processing, taking advantage of linear optical accelerators that perform operations at lower power consumption. However, to achieve efficient computing with light, it remains a challenge to create and control nonlinearity optically rather than electronically. In this study, a reservoir computing approach (RC) is investigated using a 14-mm waveguide in LiNbO3 on an insulator as an optical processor to validate the benefit of optical nonlinearity. Data are encoded on the spectrum of a femtosecond pulse, which is launched into the waveguide. The output of the waveguide is a nonlinear transform of the input, enabled by optical nonlinearities. We show experimentally that a simple digital linear classifier using the output spectrum of the waveguide increases the classification accuracy of several databases by ∼10% compared to untransformed data. In comparison, a digital neural network (NN) with tens of thousands of parameters was required to achieve similar accuracy. With the ability to reduce the number of parameters by a factor of at least 20, an integrated optical RC approach can attain a performance on a par with a digital NN.
- 1 October 2023
- Published by AIP Publishing in APL Photonics
- Vol. 8 (10)
Open AccessNonlinear optical feature generator for machine learning
Modern machine learning models use an ever-increasing number of parameters to train (175 × 109 parameters for GPT-3) with large datasets to achieve better performance. Optical computing has been rediscovered as a potential solution for large-scale data processing, taking advantage of linear optical accelerators that perform operations at lower power consumption. However, to achieve efficient computing with light, it remains a challenge to create and control nonlinearity optically rather than electronically. In this study, a reservoir computing approach (RC) is investigated using a 14-mm waveguide in LiNbO3 on an insulator as an optical processor to validate the benefit of optical nonlinearity. Data are encoded on the spectrum of a femtosecond pulse, which is launched into the waveguide. The output of the waveguide is a nonlinear transform of the input, enabled by optical nonlinearities. We show experimentally that a simple digital linear classifier using the output spectrum of the waveguide increases the classification accuracy of several databases by ∼10% compared to untransformed data. In comparison, a digital neural network (NN) with tens of thousands of parameters was required to achieve similar accuracy. With the ability to reduce the number of parameters by a factor of at least 20, an integrated optical RC approach can attain a performance on a par with a digital NN.
- Open Access
Forward-Forward Training of an Optical Neural Network
Neural networks (NN) have demonstrated remarkable capabilities in various tasks, but their computation-intensive nature demands faster and more energy-efficient hardware implementations. Optics-based platforms, using technologies such as silicon photonics and spatial light modulators, offer promising avenues for achieving this goal. However, training multiple trainable layers in tandem with these physical systems poses challenges, as they are difficult to fully characterize and describe with differentiable functions, hindering the use of error backpropagation algorithm. The recently introduced Forward-Forward Algorithm (FFA) eliminates the need for perfect characterization of the learning system and shows promise for efficient training with large numbers of programmable parameters. The FFA does not require backpropagating an error signal to update the weights, rather the weights are updated by only sending information in one direction. The local loss function for each set of trainable weights enables low-power analog hardware implementations without resorting to metaheuristic algorithms or reinforcement learning. In this paper, we present an experiment utilizing multimode nonlinear wave propagation in an optical fiber demonstrating the feasibility of the FFA approach using an optical system. The results show that incorporating optical transforms in multilayer NN architectures trained with the FFA, can lead to performance improvements, even with a relatively small number of trainable weights. The proposed method offers a new path to the challenge of training optical NNs and provides insights into leveraging physical transformations for enhancing NN performance.
Open AccessForward-Forward Training of an Optical Neural Network
Neural networks (NN) have demonstrated remarkable capabilities in various tasks, but their computation-intensive nature demands faster and more energy-efficient hardware implementations. Optics-based platforms, using technologies such as silicon photonics and spatial light modulators, offer promising avenues for achieving this goal. However, training multiple trainable layers in tandem with these physical systems poses challenges, as they are difficult to fully characterize and describe with differentiable functions, hindering the use of error backpropagation algorithm. The recently introduced Forward-Forward Algorithm (FFA) eliminates the need for perfect characterization of the learning system and shows promise for efficient training with large numbers of programmable parameters. The FFA does not require backpropagating an error signal to update the weights, rather the weights are updated by only sending information in one direction. The local loss function for each set of trainable weights enables low-power analog hardware implementations without resorting to metaheuristic algorithms or reinforcement learning. In this paper, we present an experiment utilizing multimode nonlinear wave propagation in an optical fiber demonstrating the feasibility of the FFA approach using an optical system. The results show that incorporating optical transforms in multilayer NN architectures trained with the FFA, can lead to performance improvements, even with a relatively small number of trainable weights. The proposed method offers a new path to the challenge of training optical NNs and provides insights into leveraging physical transformations for enhancing NN performance.