Topology-Aware Classifier for Class-Incremental Learning

A new classifier called Hierarchical-Cluster SOINN (HC-SOINN) is proposed to address limitations of the Nearest Class Mean (NCM) classifier in Class-Incremental Learning (CIL). NCM assumes features collapse into single points, but non-linear feature drift and insufficient training cause classes to form complex manifolds. HC-SOINN captures topological structure via a local-to-global representation. Additionally, the Structure-Topology Alignment via Residuals (STAR) method uses fine-grained pointwise trajectory tracking to deform learned topology, adapting to complex non-linear features. The paper is available on arXiv.