Optimizing photonic reservoir computers from input to output layer

Reservoir computing (RC) is a computing technique inspired by the way the human brain
processes information. While the brain relies on transient neuronal activity excited by
input sensory signals, RC rather exploits the transient behavior of an analog nonlinear
dynamical system. Using photonic components to physically implement RC has several
benefits over the conventional digital von Neumann approach, including a low-energy
consumption, high bandwidth and the possibility of high inherent parallelism of pho-
tonics. Over the years, a large variety of photonic RC systems have been proposed and
developed, all with closely related architectures and design principles. In this thesis, we
want to revisit the structure and workings of several of these photonic reservoir computers.
Putting aside any preconceptions or biases, we adapt each layer of these RC systems
with the goal to obtain considerable performance gains. The first system under study is
delay-based RC using a single-mode semiconductor laser. We revisit the type of infor-
mation encoding used in the optical injection signal at the input layer. By relying on a
phase modulation based data injection scheme we can greatly improve the performance.
At the output layer where the training is performed, we show that we can use the con-
cept of transfer learning to re-use information between different learning tasks, thereby
reducing the cost of retraining. Additionally, we demonstrate that we can replace the
traditional linear perceptron by a relatively small deep neural network, which improves
the performance of the RC system. As a second RC system, we have focused on a passive
spatially distributed RC system consisting of a network of waveguides connected via op-
tical splitters and combiners. We have investigated the incorporation of an additional
active nonlinear component into this system. Our approach involves the integration of a
single semiconductor laser in an external electro-optical delay line within the network.
Besides compensating for losses, this new RC system has a significantly increased nonlin-
ear computational capacity. Additionally, this network can also be used as a delay-based
RC approach but with the complex preprocessing procedure removed.