Recurrent Neural Network ODE Output for Classification Problems Follows the Replicator Dynamics

This letter establishes a novel relationship between a class of recurrent neural networks and certain evolutionary dynamics that emerge in the context of population games. Specifically, it is shown that the output of a recurrent neural network, in the context of classification problems, coincides with the evolution of the population state in a population game. This connection is established with dynamic payoffs and under replicator evolutionary dynamics. The connection provides insights into the neural network’s behavior from both dynamical systems and game-theoretical perspectives, aligning with recent literature that suggests that neural network outputs may resemble the Nash equilibria of suitable games. It also uncovers potential connections between the neural network classification problem and mechanism design. To illustrate our results, we present different numerical experiments in the context of classification problems.