Advanced search
1 file | 676.25 KB

Reservoir computing: a photonic neural network for information processing

Author
Organization
Abstract
At the boundaries between photonics and dynamic systems theory, we combine recent advances in neural networks with opto-electronic nonlinearities to demonstrate a new way to perform optical information processing. The concept of reservoir computing arose recently as a powerful solution to the issue of training recurrent neural networks. Indeed, it is comparable to, or even outperforms, other state of the art solutions for tasks such as speech recognition or time series prediction. As it is based on a static topology, it allows making the most of very simple physical architectures having complex nonlinear dynamics. The method is inherently robust to noise and does not require explicit programming operations. It is therefore particularly well adapted for analog realizations. Among the various implementations of the concept that have been proposed, we focus on the field of optics. Our experimental reservoir computer is based on opto-electronic technology, and can be viewed as an intermediate step towards an all optical device. Our fiber optics system is based on a nonlinear feedback loop operating at the threshold of chaos. In its present preliminary stage it is already capable of complicated tasks like modeling nonlinear systems with memory. Our aim is to demonstrate that such an analog reservoir can have performances comparable to state of the art digital implementations of Neural Networks. Furthermore, our system can in principle be operated at very high frequencies thanks to the high speed of photonic devices. Thus one could envisage targeting applications such as online information processing in broadband telecommunications.
Keywords
Information processing, reservoir computing, neural networks, machine learning, dynamical systens, photonics

Downloads

  • ReservoirComputingManuscript6.pdf
    • full text
    • |
    • open access
    • |
    • PDF
    • |
    • 676.25 KB

Citation

Please use this url to cite or link to this publication:

Chicago
Paquot, Yvan, Joni Dambre, Benjamin Schrauwen, Marc Haelterman, and Serge Massar. 2010. “Reservoir Computing: a Photonic Neural Network for Information Processing.” In Proceedings of the Society of Photo-optical Instrumentation Engineers (spie), ed. Benjamin J Eggleton, Alexander L Gaeta, and Neil GR Broderick. Vol. 7728. Bellingham, WA, USA: SPIE, the International Society for Optical Engineering.
APA
Paquot, Y., Dambre, J., Schrauwen, B., Haelterman, M., & Massar, S. (2010). Reservoir computing: a photonic neural network for information processing. In B. J. Eggleton, A. L. Gaeta, & N. G. Broderick (Eds.), PROCEEDINGS OF THE SOCIETY OF PHOTO-OPTICAL INSTRUMENTATION ENGINEERS (SPIE) (Vol. 7728). Presented at the Conference on Nonlinear Optics and Applications IV, Bellingham, WA, USA: SPIE, the International Society for Optical Engineering.
Vancouver
1.
Paquot Y, Dambre J, Schrauwen B, Haelterman M, Massar S. Reservoir computing: a photonic neural network for information processing. In: Eggleton BJ, Gaeta AL, Broderick NG, editors. PROCEEDINGS OF THE SOCIETY OF PHOTO-OPTICAL INSTRUMENTATION ENGINEERS (SPIE). Bellingham, WA, USA: SPIE, the International Society for Optical Engineering; 2010.
MLA
Paquot, Yvan, Joni Dambre, Benjamin Schrauwen, et al. “Reservoir Computing: a Photonic Neural Network for Information Processing.” Proceedings of the Society of Photo-optical Instrumentation Engineers (spie). Ed. Benjamin J Eggleton, Alexander L Gaeta, & Neil GR Broderick. Vol. 7728. Bellingham, WA, USA: SPIE, the International Society for Optical Engineering, 2010. Print.
@inproceedings{1070018,
  abstract     = {At the boundaries between photonics and dynamic systems theory, we combine recent advances in neural networks with opto-electronic nonlinearities to demonstrate a new way to perform optical information processing. The concept of reservoir computing arose recently as a powerful solution to the issue of training recurrent neural networks. Indeed, it is comparable to, or even outperforms, other state of the art solutions for tasks such as speech recognition or time series prediction. As it is based on a static topology, it allows making the most of very simple physical architectures having complex nonlinear dynamics. The method is inherently robust to noise and does not require explicit programming operations. It is therefore particularly well adapted for analog realizations. Among the various implementations of the concept that have been proposed, we focus on the field of optics. Our experimental reservoir computer is based on opto-electronic technology, and can be viewed as an intermediate step towards an all optical device. Our fiber optics system is based on a nonlinear feedback loop operating at the threshold of chaos. In its present preliminary stage it is already capable of complicated tasks like modeling nonlinear systems with memory. Our aim is to demonstrate that such an analog reservoir can have performances comparable to state of the art digital implementations of Neural Networks. Furthermore, our system can in principle be operated at very high frequencies thanks to the high speed of photonic devices. Thus one could envisage targeting applications such as online information processing in broadband telecommunications.},
  articleno    = {77280B},
  author       = {Paquot, Yvan and Dambre, Joni and Schrauwen, Benjamin and Haelterman, Marc and Massar, Serge},
  booktitle    = {PROCEEDINGS OF THE SOCIETY OF PHOTO-OPTICAL INSTRUMENTATION ENGINEERS (SPIE)},
  editor       = {Eggleton, Benjamin J and Gaeta, Alexander L and Broderick, Neil GR},
  isbn         = {9780819482013},
  issn         = {0277-786X},
  keyword      = {Information processing,reservoir computing,neural networks,machine learning,dynamical systens,photonics},
  language     = {eng},
  location     = {Brussels, Belgium},
  pages        = {12},
  publisher    = {SPIE, the International Society for Optical Engineering},
  title        = {Reservoir computing: a photonic neural network for information processing},
  url          = {http://dx.doi.org/10.1117/12.854050},
  volume       = {7728},
  year         = {2010},
}

Altmetric
View in Altmetric
Web of Science
Times cited: