Finite-Time Convergent Distributed Cooperative Learning Algorithm for Data Approximation

This paper aims to solve the distributed cooperative learning (DCL) problem over networks, such as data approximation, where each node only has access to local information which is produced by the same unknown pattern ( map or function). Different from the traditional centralized learning (CL) scheme, DCL scheme needs all nodes in the network cooperatively learn the unknown pattern by exchanging its own learned information with their neighbors. In order to share learned information of each node, a novel finite-time convergent DCL algorithm via High-Order Neural Networks (IIONN) over undirected network with fixed topologies is developed. The numerical experiment and rigorous theoretical analysis show that not only the proposed algorithm owns high learning ability, but also owns high rate of convergence.