Higher-order interactions, adaptivity, and phase transitions in a novel reservoir computing model.

We propose a novel reservoir neural network model that incorporates key properties of brain neural ensembles, including adaptivity, higher-order interactions among units, and the presence of a phase transition, which allows "edge-of-chaos computations." The network's performance was evaluated on benchmark machine learning tasks, such as reproducing multidimensional periodic patterns and predicting the dynamics of the chaotic Lorenz attractor. Our findings indicate that interelement couplings primarily contribute to generating the target output. Furthermore, we demonstrate that a new phase transition occurs after learning, such that the dynamics of the phases become different from the initial.