Dynamic Graph Convolution Network for Multi-Label Class Incremental Learning

While single-label class incremental learning (CIL) has seen significant progress, multi-label class incremental learning (MLCIL) remains relatively unexplored. In MLCIL, newly introduced samples are exclusively labeled with novel classes, lacking annotations for previously learned ones. To address the resulting missing label information, generating pseudo-labels has become a common approach. In CIL, models gradually forget what they have already learned, which is called catastrophic forgetting. Similar to CIL, MLCIL is also susceptible to catastrophic forgetting. To tackle these challenges, we propose an Incremental Multi-Label Graph Convolution Network (IMLGCN) for MLCIL, which is constructed by the Dynamic Graph Convolution Graph (D-GConv). The D-GConv constructs topological relationships between new and learned classes within the current session, which will effectively extract the information from a global view. Extensive experiments on standard datasets demonstrate the superior performance of our proposed model.