Global and Local Graph Convolution Network for Multi-Label Class Incremental Learning

Despite significant progresses in single-label class incremental learning (CIL), its multi-label counterpart remains relatively understudied which is called multi-label class incremental learning (MLCIL). In MLCIL, newly introduced samples are exclusively labeled with new classes, necessitating heavy reliance on pseudo-label generation to compensate for missing annotations. Simply adapting CIL methods to MLCIL is ineffective due to the challenges of catastrophic forgetting and lack of learned labels. To address these issues, we propose Global and Local Graph Convolution Network (GL-GCN), incorporating Global Graph Convolution Network (G-GConv) and Local Graph Convolution Network (L-GConv) modules. GGConv captures inter-session connections, while L-GConv preserves intra-session class relationships for knowledge retention. Additionally, asymmetric loss is introduced to mitigate class imbalance. Experimental results demonstrate our method’s superiority over existing approaches.