Abstract
Online Continual Learning (CL) solves the problem of learning the ever-emerging new classification tasks from a continuous data stream. Unlike its offline counterpart, in online CL, the training data can only be seen once. Most existing online CL research regards catastrophic forgetting (i.e., model stability) as almost the only challenge. In this paper, we argue that the model's capability to acquire new knowledge (i.e., model plasticity) is another challenge in online CL. While replay-based strategies have been shown to be effective in alleviating catastrophic forgetting, there is a notable gap in research attention toward improving model plasticity. To this end, we propose Collaborative Continual Learning (CCL), a collaborative learning based strategy to improve the model's capability in acquiring new concepts. Additionally, we introduce Distillation Chain (DC), a collaborative learning scheme to boost the training of the models. We adapt CCL-DC to existing representative online CL works. Extensive experiments demonstrate that even if the learners are well-trained with state-of-the-art online CL methods, our strategy can still improve model plasticity dramatically, and thereby improve the overall performance by a large margin.
Introduction
In Online Continual Learning (CL), an AI model encounters a stream of data and must learn new tasks sequentially. Unlike "offline" learning, the model only sees each data point once. Historically, researchers have focused on stability: ensuring the model doesn't forget Task A while learning Task B. However, this paper identifies a "plasticity gap." Because the model only gets one look at new data, it often fails to learn the new task effectively in the first place.
As shown in Figure 1, there is a significant drop in Learning Accuracy when moving from offline to online settings. If a model does not learn a task well initially, there is little knowledge to preserve. This research shifts the focus toward improving plasticity—the efficiency and speed of acquiring new knowledge on the fly.
Methodology: The CCL-DC Framework
The authors propose a two-part framework called CCL-DC to address the plasticity bottleneck.
Collaborative Continual Learning (CCL)
Instead of training a single model, CCL trains two "peer" models simultaneously. These models engage in peer-teaching by regularizing each other's predictions. This parallel approach facilitates faster convergence on new data, which is critical in the single-epoch constraint of online CL.
Distillation Chain (DC)
Models in online CL often become overconfident too quickly, hindering generalization. The Distillation Chain addresses this by creating a sequence of images with increasing geometric distortion and augmentation (an "Augmentation Chain" from easy to hard). The model distills knowledge from the "harder" (less confident) samples to the "easier" (more confident) ones. This acts as a form of entropy regularization, preventing the model from over-fitting to a single view of the data and maintaining the plasticity needed for better representation learning.
Experimental Results
The framework was evaluated on standard image classification benchmarks including CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-100. The authors introduced two refined metrics: Learning Accuracy (LA) to quantify plasticity and Relative Forgetting (RF) to assess stability fairly across models with different initial learning capabilities.
The CCL-DC framework was integrated into six state-of-the-art online CL methods, such as Experience Replay (ER), DER++, and OCM. Significant improvements were observed across the board:
- Accuracy Boost: On Tiny-ImageNet, the baseline ER method improved from 10.82% to 16.56% average accuracy.
- Plasticity Gains: The Learning Accuracy (LA) for ER on CIFAR-100 increased from 53.77% to 72.38%, demonstrating a much higher capacity for picking up new concepts.
- Robustness: The method consistently improved performance regardless of the memory buffer size, indicating it is a robust architectural enhancement.
Conclusion
This work highlights that learning fast is just as important as not forgetting in the context of online continual learning. By introducing CCL-DC, the authors provide a mechanism to boost model plasticity through collaborative learning and structured distillation. While the approach increases computational cost—roughly doubling training time—the gains in accuracy and learning efficiency are substantial. Future work may explore making this collaborative process more efficient for resource-constrained environments.
Summary updated