Learn the Time to Learn: Replay Scheduling for Continual Learning

Marcus Klasson, Hedvig Kjellstrom, Cheng Zhang

Keywords: Continual Learning, Replay Memory, Task Incremental Learning

Abstract: Replay-based continual learning has shown to be successful in mitigating catastrophic forgetting. Most previous works focus on increasing the sample quality in the commonly small replay memory. However, in many real-world applications, replay memories would be limited by constraints on processing time rather than storage capacity as most organizations do store all historical data in the cloud. Inspired by human learning, we illustrate that scheduling over which tasks to revisit is critical to the final performance with finite memory resources. To this end, we propose to learn the time to learn for a continual learning system, in which we learn schedules over which tasks to replay at different times using Monte Carlo tree search. We perform extensive evaluation and show that our method can learn replay schedules that significantly improve final performance across all tasks than baselines without considering the scheduling. Furthermore, our method can be combined with any other memory selection methods leading to consistently improved performance. Our results indicate that the learned schedules are also consistent with human learning insights.