Information-theoretic Online Memory Selection for Continual Learning

A challenging problem in task-free continual learning is the online selection of a representative replay memory from data streams. In this work, we investigate the online memory selection problem from an information-theoretic perspective. To gather the most information, we propose the surprise and the learnability criteria to pick informative points and to avoid outliers. We present a Bayesian model to compute the criteria efficiently by exploiting rank-one matrix structures. We demonstrate that these criteria encourage selecting informative points in a greedy algorithm for online memory selection. Furthermore, by identifying the importance of the timing to update the memory, we introduce a stochastic informationtheoretic reservoir sampler (InfoRS), which conducts sampling among selective points with high information. Compared to reservoir sampling, InfoRS demonstrates improved robustness against data imbalance. Continual learning (Robins, 1995; Goodfellow et al., 2013; Kirkpatrick et al., 2017) aims at training models through a non-stationary data stream without catastrophic forgetting of past experiences. Specifically, replay-based methods (Lopez-Paz & Ranzato, 2017; Rebuffi et al., 2017; Rolnick et al., 2019) tackle the continual learning problem by keeping a replay memory for rehearsals over the past data. Given the limited memory budget, selecting a representative memory becomes critical. The majority of existing approaches focus on task-based continual learning and update the memory based on the given task boundaries. Since the requirement for task boundaries is usually not realistic, general continual learning (GCL) (Aljundi et al., 2019a; Delange et al., 2021; Buzzega et al., 2020) has received increasing attention, which assumes that the agent observes the streaming data in an online fashion without knowing task boundaries. GCL makes the online memory selection more challenging since one needs to update the memory in each iteration based only on instant observations. So, successful memory management for GCL needs to be both efficient and effective.