Towards a Unified Lifelong Learning Framework
NeurIPS 2020 Workshop on Pre-registration in Machine Learning, PMLR 148:221-235, 2021.
Humans can learn a variety of concepts and skills incrementally over the course of their lives while exhibiting many desirable properties, such as continual learning without forgetting, forward transfer of knowledge, and learning a new concept with few examples. However, most previous approaches to efficient lifelong learning demonstrate only subsets of these properties, often by different complex mechanisms. In this preregistration submission, we propose to study the effectiveness of a unified lifelong learning framework designed to achieve many of these properties through one central mechanism. We describe this consolidation-based approach and propose experimental protocols to benchmark it on several skills, using grid searches over hyperparameters to better understand the framework.