Multi-Model Adaptive Ensemble Method for Fewshot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) aims to enable models to continuously learn novel classes with limited samples, yet balancing feature stability and plasticity during incremental learning remains a critical challenge. To address the difficulty in capturing task-specific details during incremental sessions in FSCIL, we propose a multi-model adaptive ensemble method (MMAE). The framework enhances the diversity of feature representations through a multi-model architecture, employs an R-dropout regularization strategy to suppress interference from redundant feature information, and constructs a gated network to dynamically focus on the contribution weights of sub-models. This framework achieves precise capture of taskspecific knowledge and optimizes classification decisions in incremental sessions. Experiments on three widely-used few-shot class-incremental learning datasets validate the effectiveness of the MMAE model in this task.