Information processing with photonic hardware neural networks

This thesis presents numerical and experimental investigations that validate the use of photonic systems as information processors. For that, the complex temporal dynamical properties of semiconductor lasers subject to delayed optical feedback and optical injection have been investigated.

The first part of the thesis presents the experimental setup used for information processing. The different parts that form the time delay reservoir computer (TDRC) are explored by gradually investigating the contribution of different disturbances to the laser emission. A laser subjected to feedback, then injection, then feedback and injection, and lastly, also modulated injection. Graphs regarding the dynamics of the system are obtained utilizing equipment capable of high bandwidth RF and optical spectrum measurements.

The second part of the thesis leverages the master-slave configuration present in our TDRC to generate system responses with speeds up to tens of gigahertz. The response bandwidth is increased by employing strong optical injection of the drive laser into reservoir one. This bandwidth enhancement is proved to speed up the photonic TDRC computation without sacrificing performance. It is proved firstly through numerical results and later demonstrated experimentally. Two main conclusions are drawn from the study. First, a wider region of better performance is achieved for strong optical condition. This is important since a robust operation of the TDRC as a processing system is critical for establishing this technique as a credible candidate for ultrafast analog signal processing. Second, a faster system response allows for faster information processing. A nonlinear transfer function that generates transient responses at tens of gigahertz speed is demonstrated for different tasks.

Efficient data recovery in fiber transmission systems is demonstrated in the third part of the thesis. Photonic processing schemes (TDRC and extreme learning machines, ELM) are compared to digital ones (Kramers-Kronig receiver) for the post-processing after an experimental transmission link. For the ELM approach, an improvement on the data recovery performance is achieved in comparison to that of a digital one for certain optical signal-to-noise ratio (OSNR). The advantages and disadvantages of using hardware photonic systems for data recovery are also discussed.

Alternative photonic processing architecture is presented in the last part of the thesis, where wavelength division multiplexing is used instead of the time-multiplexing technique in the TDRC approach. An experimental setup mimicking a tapped delay neural network is implemented experimentally and tested against digital to analog conversion and header recognition tasks.

Altogether, the thesis presents results that highlight the future of photonics in the field of information processing. The ideas presented here have the potential to be used for commercial applications.