Lifelong Learning

Lifelong learning is the open-ended, cumulative acquisition of knowledge and skills over an extended lifetime, without predefined task boundaries or fixed curricula. In machine learning, it is often used interchangeably with continual learning, but the two concepts diverge at a crucial point: continual learning is typically evaluated on discrete, labeled task sequences with measurable forgetting, while lifelong learning encompasses the messier reality of unstructured experience — learning from conversation, exploration, failure, and social interaction without clear task delineations. The distinction matters because the benchmarks of continual learning may not capture the problem that lifelong learning actually poses.

Biological lifelong learning is the norm, not the exception. Humans do not learn language, then music, then physics as separate tasks with separate datasets. We learn continuously from overlapping, unlabeled experience, and our knowledge structures are constantly reorganized. The gap between artificial continual learning benchmarks and biological lifelong learning suggests that the field may be optimizing for the wrong metric: not forgetting on previous tasks, but the capacity to recombine old knowledge into new competencies that no one, including the learner, knew would be needed.