Mitigating the Stability-Plasticity Trade-off in Neural Networks via Shared Extractors in Class-Incremental Learning

Humans learn new tasks without forgetting, but neural networks suffer catastrophic forgetting when trained sequentially. Dynamic expandable networks attempt to address this by assigning each task its own feature extractor and freezing previous ones to preserve past knowledge. While effective for retaining old tasks, this design leads to rapid parameter growth, and frozen extractors never adapt to future data, often producing irrelevant features that degrade later performance. To overcome these limitations, we propose Task-Sharing Distillation (TSD), which reduces the number of extractors by allowing multiple tasks to share one extractor and consolidating them through distillation. We study two strategies: (1) Grouped Rolling Consolidation, which groups consecutive tasks and consolidates them into a shared extractor, and (2) Fixed-Size Pool with Similarity-Based Consolidation, where new tasks are merged into the most compatible extractor in a fixed pool according to prototype similarity. Experiments on CIFAR-100 and ImageNet-100 show that TSD maintains strong performance across tasks, demonstrating that careful feature sharing is more effective than simply adding more extractors. The idea also has potential relevance for large language models, where continual and multi-domain adaptation face similar challenges of parameter growth and redundancy.