FlexGP : a scalable system for factored learning in the cloud

[摘要] This work presents FlexGP, a new system designed for scalable machine learning in the cloud. FlexGP presents a learner-agnostic, data-parallel approach to cloud-based distributed learning using existing single-machine algorithms, without any dependence on distributed file systems or shared memory between instances. We design and implement asynchronous and decentralized launch and peer discovery protocols to start and configure a distributed network of learners. Through a unique process of factoring the data and parameters across the learners, FlexGP ensures this network consists of heterogeneous learners producing diverse models. These models are then filtered and fused to produce a meta-model for prediction. Using a thoughtfully designed test framework, FlexGP is run on a real-world regression problem from a large database. The results demonstrate the reliability and robustness of the system, even when learning from very little training data and multiple factorings, and demonstrate FlexGP as a vital tool to effectively leverage the cloud for machine learning tasks.

[发布机构] Massachusetts Institute of Technology