Dynamically Anchored Prompting for Task-Imbalanced Continual Learning

TL;DR

This work addresses task-imbalanced continual learning (TICL), where data across tasks arrives with nonuniform sizes, complicating the balance between retaining old knowledge and learning new tasks. It introduces Dynamically Anchored Prompting (DAP), a prompt-based method that maintains a single general prompt regularized by a boosting anchor and a stabilizing anchor, enabling rehearsal-free learning across a shifting task stream. DAP employs a two-phase in-task training scheme and a dynamic stability-plasticity factor $\\lambda$ to adapt the emphasis on plasticity or stability based on current task size, with online updates to the stabilizing anchor. Empirical results on TICL benchmarks (CIFAR-100 and ImageNet-R) show substantial improvements ($4.5\%$ to $15\%$) over state-of-the-art methods, validating a memory-efficient and effective approach for real-world, imbalanced continual learning applications.

Abstract

Existing continual learning literature relies heavily on a strong assumption that tasks arrive with a balanced data stream, which is often unrealistic in real-world applications. In this work, we explore task-imbalanced continual learning (TICL) scenarios where the distribution of task data is non-uniform across the whole learning process. We find that imbalanced tasks significantly challenge the capability of models to control the trade-off between stability and plasticity from the perspective of recent prompt-based continual learning methods. On top of the above finding, we propose Dynamically Anchored Prompting (DAP), a prompt-based method that only maintains a single general prompt to adapt to the shifts within a task stream dynamically. This general prompt is regularized in the prompt space with two specifically designed prompt anchors, called boosting anchor and stabilizing anchor, to balance stability and plasticity in TICL. Remarkably, DAP achieves this balance by only storing a prompt across the data stream, therefore offering a substantial advantage in rehearsal-free CL. Extensive experiments demonstrate that the proposed DAP results in 4.5% to 15% absolute improvements over state-of-the-art methods on benchmarks under task-imbalanced settings. Our code is available at https://github.com/chenxing6666/DAP

Figures (6)

• Figure 1: An illustration of the scenario of task-imbalanced continual learning (TICL). (a) Three cases in TICL compared to ordinary task-balanced CL. (b) Performance degradation of DualPrompt on TICL CIFAR-100. The number of balanced and imbalanced tasks is ensured to be the same.
• Figure 2: Stability
• Figure 3: Plasticity
• Figure 5: Overview of the proposed dynamic anchored prompting (DAP) for task-imbalanced continual learning. Task-imbalanced training data stream represents the sequential arrival of data from Task 0 to Task t. There are two phases for each task. In in-task phase 1, the initialized task-specific prompt learns knowledge related to the current task, serving as a boosting anchor. In in-task phase 2, the general prompt is trained, with the assistance of boosting anchor to ensure plasticity. Meanwhile, the centers of all previously learned task-specific prompts serve as stabilizing anchors, emphasizing stability. It is worth noting that the boosting anchors of past tasks are not stored as the stabilizing anchor is updated in an online manner.
• Figure 6: Performance comparison between DAP and DualPrompt with linear probe.