Evolving reservoir computers reveal bidirectional coupling between predictive power and emergent dynamics.

Abstract

Biological neural networks perform complex computations to predict their environment, far exceeding the capabilities of individual neurons. Here, we argue that understanding these computations requires considering <i>emergent</i> dynamics-dynamics that make the whole system "more than the sum of its parts." We examine the relationship between prediction performance and emergence by leveraging quantitative metrics of emergence and modeling environmental time-series prediction within a bio-inspired computational framework called reservoir computing. Notably, three key results reveal a robust bidirectional coupling between prediction performance and emergence: (1) optimizing hyperparameters for performance enhances emergent dynamics, and vice versa; (2) emergent dynamics serve as a highly sufficient and often also necessary condition for prediction success in most environments; and (3) training with larger datasets results in stronger emergent dynamics, encoding task-relevant information. These findings emphasize the importance of emergence-based approaches for studying neural networks-biological or artificial-as they enable network-level insights, complementing traditional single-neuron-based analyses.