Conception and realization of optical ''Reservoir Computing'' implementations of neural networks
Problem statements
One of the challenges faced by the telecommunications industry is to find ways to process information carried by optical telecommunications cables all-optically, i.e., without the need to convert it to the electric domain. One possible way of achieving this goal is to resort to simple optical analogue computers. This would have great advantages in terms of speed and power consumption. In this project we investigate a recently proposed approach called 'Reservoir Computing' to build an all-optical processor, dedicated to solve simple, specific tasks. In it, we seek ways to leverage the inherent analogue processing capabilities of non-linear dynamic systems. The Reservoir Computing approach has already demonstrated a promising capacity for signal classification (for instance wireless or optical channel equalization, speech recognition) and signal forecasting (radar signal forecasting, chaotic time series prediction etc…).
Our current research also extends beyond the aforementioned applications. For one, we investigate ways to apply the reservoir paradigm to the control of so-called soft robots, robots that no longer consist of rigid joints and gearboxes, but rather soft and flexible materials, which offers great advantages in terms of safety and deployability. Another line of research is to extend the application domain of optical processors by going beyond the Reservoir Computing paradigm, and to consider them as physical manifestations of neural networks, entities that can be trained to solve a certain task. We study ways to implement the training process itself optically, potentially offering great benefits in terms of speed and scalability.
Contributions
Based on our knowledge of optical nonlinear systems, we investigate several possible implementations of optical reservoir computers. We already propose optoelectronic and optical implementations, from which we intend to construct completely analogue and standalone reservoir computers.
Scientific partners
- Wireless communication group, ULB (Pr. Philippe De Doncker, and Pr. François Horlin)
- Reservoir lab, Ghent University (Pr Joni Dambre)
- Photonic research group, Ghent University (Pr Peter Bienstman)
Research funding
- Interuniversity Attraction Pole program of the Belgian Science Policy Office under grant IAP P7-35 photonics@be
- Action de Recherche Concertée (ARC) : « Optical reservoir computing and application to high speed communications »
Contact
Campus du Solbosch - CP 194/05
Avenue F.D. Roosevelt, 50
1050 Bruxelles
Campus de la Plaine - CP 224
Boulevard du Triomphe, ACC.2
1050 Bruxelles