Incentivizing Participation to Distributed Neural Network Training

Abstract

During the last years a vast number of online sensors continuously generate data that can be utilized to create novel deep learning applications. Training very large models requires enormous processing power; thus, the evident way to follow is to lease the power of a corporate data center. But the diffusion of Artificial Intelligence to an always increasing number of human activities, constantly attracts new researchers who wish to train and test their models. Our work on LEARNAE is a proposal for a purely distributed neural network training, based on a peer-to-peer and permissionless architecture. LEARNAE allows individual researchers to join forces, in order to collaboratively train a model. The process utilizes modern Distributed Ledger Technology and it is fully democratized, prioritizing decentralization, fault tolerance and privacy. In this paper we add another piece to the puzzle: A method for incentivizing peers to participate to the training swarm, even if they don’t have any interest in the produced neural network. This is achieved by embedding a reward subsystem to LEARNAE; thus, peers who contribute to teamwork can receive a proportional digital payment.