Clique-based topology enhances reservoir computing performance

Reservoir computing represents a robust machine-learning approach displaying high performance in predicting the behavior of natural and artificial complex systems. However, its full potential has been so far hindered by the limitations of traditional architectures. Here, we give the practical demonstration of the fundamental role that the specific network topology may play in machine-learning systems, and we propose a novel reservoir computer whose hidden layer comprises nodes that form cliques of varying sizes. The result is a reservoir with far more advanced data representation and processing capabilities, which consistently outperforms conventional architectures in tasks including predicting stochastic neuron dynamics, modeling the evolution of chaotic systems, and reconstructing electroencephalogram signals. Furthermore, our approach addresses a pivotal challenge in the physical implementation of reservoirs: it mitigates the sensitivity to the choice of specific topologies, improving generality and facilitating the deployment of more compact and efficient architectures. Through a systematic analysis, we elucidate the functional significance of each topological element and offer practical design strategies.