One of the main characteristics in many real-world big data scenarios is their distributed nature. In a machine learning context, distributed data, together with the requirements of preserving privacy and scaling up to large networks, brings the challenge of designing fully decentralized training protocols. In this paper, we explore the problem of distributed learning when the features of every pattern are available throughout multiple agents (as is happening, for example, in a distributed database scenario). We propose an algorithm for a particular class of neural networks, known as Random Vector Functional-Link (RVFL), which is based on the Alternating Direction Method of Multipliers optimization algorithm. The proposed algorithm allows to learn an RVFL network from multiple distributed data sources, while restricting communication to the unique operation of computing a distributed average. Our experimental simulations show that the algorithm is able to achieve a generalization accuracy comparable to a fully centralized solution, while at the same time being extremely efficient.
展开▼
