GRASP: A Rehearsal Policy for Efficient Online Continual Learning

TL;DR

GRASP addresses rehearsal efficiency in online continual learning by delivering a curriculum-based sample selection that starts with easy samples and progressively includes harder ones based on distance to class prototypes. It is computationally lightweight and scalable, integrating with multiple rehearsal CL systems and delivering strong performance on large-scale vision and NLP benchmarks. Empirically, GRASP improves final and average accuracy on ImageNet-1K, reduces updates by about 40%, and halves training time in many settings, while remaining robust across IID and long-tailed distributions. This work provides a practical, scalable mechanism to balance stability and plasticity in continual learning with constrained compute.

Abstract

Continual learning (CL) in deep neural networks (DNNs) involves incrementally accumulating knowledge in a DNN from a growing data stream. A major challenge in CL is that non-stationary data streams cause catastrophic forgetting of previously learned abilities. A popular solution is rehearsal: storing past observations in a buffer and then sampling the buffer to update the DNN. Uniform sampling in a class-balanced manner is highly effective, and better sample selection policies have been elusive. Here, we propose a new sample selection policy called GRASP that selects the most prototypical (easy) samples first and then gradually selects less prototypical (harder) examples. GRASP has little additional compute or memory overhead compared to uniform selection, enabling it to scale to large datasets. Compared to 17 other rehearsal policies, GRASP achieves higher accuracy in CL experiments on ImageNet. Compared to uniform balanced sampling, GRASP achieves the same performance with 40% fewer updates. We also show that GRASP is effective for CL on five text classification datasets.

Figures (11)

• Figure 1: GRASP achieves the best accuracy of the popular uniform balanced policy while requiring $40\%$ fewer gradient descent updates and $36\%$ less training time for CIL with SIESTA on ImageNet-1K.
• Figure 2: Unlike GRASP, existing state-of-the-art methods require significantly longer training time than uniform.
• Figure 3: (a) The ImageNet-1K learning curves of GRASP and uniform balanced policies in CIL using SIESTA. (b and c) The final accuracy of various rehearsal policies in CIL on ImageNet-300 using SIESTA and latent rehearsal.
• Figure 4: The representation drift of old classes while learning new classes in a stream of two tasks denoted by background colors. GRASP reduces representation drift.
• Figure 5: Overview of GRASP and Random Rehearsal Policies. Class mean is denoted by star. Selected samples are indicated by red circle.