How to Intelligently Distribute Training Data to Multiple Compute Nodes: Distributed Machine Learning via Submodular Partitioning

In this paper we investigate the problem of training data partitioning for parallel learning of statistical models. Motivated by [10], we utilize submodular functions to model the utility of data subsets for training machine learning classifiers and formulate this problem mathematically as submodular partitioning. We introduce a simple and scalable greedy algorithm that near-optimally solves the submodular partitioning problem. We empirically demonstrate the efficacy of the proposed algorithm to obtain data partitioning for distributed optimization of convex and deep neural network objectives. Empirical evidences suggest that the intelligent data partitioning produced by the proposed framework leads to faster convergence in the case of distributed convex optimization, and better resulting models in the case of parallel neural network training.