Continual learning requires models to integrate new skills while preserving existing knowledge, yet most modular approaches either over-expand their architectures or fail to preserve functional reuse across tasks. We propose the Pooling-Gated Modular Networks (PGMN), a scalable framework that achieves controlled dynamic expansion through a task-conditioned pooling gate and a fixed-capacity module pool. The pooling-gated mechanism produces low-dimensional routing weights that promote consistent reuse of previously acquired modules, while new modules are instantiated only when the model cannot sufficiently represent a new task. To further stabilize learning under sequential updates, we introduce the Module Weight Regularizer (MWR), which constrains gating shifts and encourages principled reuse of frozen modules, reducing representational drift and enabling effective forward transfer. Evaluations on the Continual Transfer Learning (CTrL) benchmark, including challenging long-horizon task streams, show that PGMN significantly improves task recognition accuracy, reduces forgetting, and maintains strict control over architectural growth compared to existing modular continual-learning baselines. These results demonstrate that PGMN provides an effective balance between plasticity, stability, and scalability in realistic continual-learning scenarios.
Fang et al. (Sun,) studied this question.