Reservoir Computing Based Cryptography and Exploration of the Limits of Multifunctionality in NG-RC

Abstract

Reservoir computing has become the state-of-the-art machine learning algorithm for predicting nonlinear and chaotic dynamics. It features excellent speed and less required training data compared to other deep learning methods. The first part of this thesis makes use of the algorithm’s speed aspect. A new encryption algorithm is developed, which outperforms a previous reservoir computing based encryption algorithm by a factor of 1000 in terms of encryption speed. Reservoir computing was also successfully applied to simulate biological neural functions. One of these functions is learning multiple tasks with the identical network structure simultaneously, i.e. the ability to be multifunctional. In reservoir computing, the intrinsic network structure is not changed during multifunctional processing, resembling its biological counterpart. The next generation of reservoir computing (NG-RC) was recently introduced, featuring improved performance. Therefore, the functioning of the reservoir network is replaced by polynomial multiplications of time-shifted input variables. The second part of this thesis explores the limits of multifunctionality in NG-RC. The architecture of the algorithm creates high interpretability of multifunctional behavior. This opens a new perspective on multifunctionality and allows such behavior to be analyzed by learned governing equations.