On-device, online continual learning for the real world

Abstract

Humans possess the ability to learn a large number of tasks by accumulating knowledge and skills over time. Building a system resembling human learning abilities is a deep-rooted desire since sustainable learning over a long-term period is essential for general artificial intelligence. In light of this need, continual learning (CL), or lifelong learning, tackles a learning scenario where a model continuously learns over a sequence of tasks within a broad research area. For my long-term research goal, I aim to encompass broad research fields to understand humans and impact our real lives through sustainable on-device AI systems with multiple agents. Embedded machines persistently learn human users’ experiences, formulated to non-stationary online egocentric video streams. While multiple agents learn their local experiences separately, they can communicate with each other to expand and evolve their knowledge hosted by the server. The server contributes to stacking and merging the local knowledge on various problems and re-distributing them to participating agents on demand. However, this learning paradigm raises several crucial challenges, First, on-device AI has limited computational and memory budgets according to the hardware and difficulties of local sequential tasks. Second, the agent should adaptively transfer the relevant knowledge from other devices’ experience. Further, we aim to train on real-world data in an online manner, and the real-world data contain a large amount of redundant, imbalanced, and noisy, as well as unlabeled instances. That is, the model should cope with not refined and unannotated training data. And this thesis proposes methods from multiple perspectives for solving these challenges of sustainable on-device, online continual learning for the real world.