Simple, Efficient and Convenient Decentralized Multi-task Learning for Neural Networks

Abstract

Machine learning requires large amounts of data, which is increasingly distributed over many systems (user devices, independent storage systems). Unfortunately aggregating this data in one site for learning is not always practical, either because of network costs or privacy concerns. Decentralized machine learning holds the potential to address these concerns, but unfortunately, most approaches proposed so far for distributed learning with neural network are mono-task, and do not transfer easily to multi-tasks problems, for which users seek to solve related but distinct learning tasks and the few existing multi-task approaches have serious limitations. In this paper, we propose a novel learning method for neural networks that is decentralized, multi-task, and keeps users' data local. Our approach works with different learning algorithms, on various types of neural networks. We formally analyze the convergence of our method, and we evaluate its efficiency in different situations on various kind of neural networks, with different learning algorithms, thus demonstrating its benefits in terms of learning quality and convergence. I IntroductionA critical requirement for machine learning is training data. In some cases, a great amount of data is available from different (potential) users of the system, but simply aggregating this data and using it for training is not always practical. The data might for instance be large and extremely distributed and collecting it may incur significant communication cost. Users may also be unwilling to share their data due to its potential sensitive nature, as is the case with private conversations, browsing histories, or health-related data [Che+17].To address these issues, several works have proposed to share model-related information (such as gradients or model coefficients) rather than raw data [Rec+11; Kon+16; Bre+17]. These approaches are however typically mono-task, in the sense that all users are assumed to be solving the same ML task. Unfortunately, in a distributed setting, the problems that users want to solve may not be perfectly identical. Let us consider the example of speech recognition being performed on mobile device. At the user level, each has a different voice, and so the different devices do not perform exactly the same task. At the level of a country or region, language variants also impose variants between tasks. For example users from Quebec can clearly be separated from those from France. A decentralized or federated learning platform should therefore accommodate for both types of differences between tasks: at the level of single users and at the level of groups of users.Multitask learning has been widely studied in a centralized setting [Rud17]. Some distributed solutions exist, but they are typically limited to either linear [OHJ12] or convex [Smi+16; Bel+18; ZBT19] optimization problems. As a result, they are typically not applicable to neural networks, in spite of their high versatility and general success in solving a broad range of machine-learning 1/39 D...