Multi-Agent Distributed Lifelong Learning for Collective Knowledge Acquisition.

Lifelong machine learning methods acquire knowledge over a series ofconsecutive tasks, continually building upon their experience. Current lifelonglearning algorithms rely upon a single learning agent that has centralizedaccess to all data. In this paper, we extend the idea of lifelong learning froma single agent to a network of multiple agents that collectively learn a seriesof tasks. Each agent faces some (potentially unique) set of tasks; the key ideais that knowledge learned from these tasks may benefit other agents trying tolearn different (but related) tasks. Our Collective Lifelong Learning Algorithm(CoLLA) provides an efficient way for a network of agents to share theirlearned knowledge in a distributed and decentralized manner, while preservingthe privacy of the locally observed data. Note that a decentralized scheme is asubclass of distributed algorithms where a central server does not exist and inaddition to data, computations are also distributed among the agents. Weprovide theoretical guarantees for robust performance of the algorithm andempirically demonstrate that CoLLA outperforms existing approaches fordistributed multi-task learning on a variety of data sets.