Human intelligence is remarkable in terms of generalization. A human has outstanding capability to generalize prior knowledge to a novel task with somewhat different goal or context, using little data and minimal effort. This strong generalization capability is one of the major differences between human intelligence and today’s mainstream AI systems which focus on narrow AI.
In this project, we propose, formalize, explore new approaches and create new benchmark datasets for Heterogeneous Meta-Learning with the goal of achieving generalization performance closer to human intelligence. We have identified critical research gaps pertaining to meta-learning of heterogeneous tasks, and will develop new methods to address these gaps in the project.