A new preprint on arXiv, LSA-PAST, tackles class-incremental learning (CIL) in settings where visual concepts are compositional—that is, formed from combinations of simpler primitives. Standard CIL lets models recognize new classes while preserving old knowledge, but the authors argue that many real-world problems require handling such compositional structures.
The proposed method, Layer-Specific Adaptation and Prototype Transport, appears designed to address this by adapting different network layers in a targeted way and transporting prototypes to support incremental learning. The abstract emphasizes the gap between conventional CIL and compositional scenarios, suggesting that LSA-PAST is a step toward more flexible continual learning.
As an arXiv announcement, the details are limited to the abstract, so no experimental results or comparisons are available in this source. The work is notable for focusing on compositional rather than flat class incremental learning, a distinction that matters for applications where concepts share underlying primitives.