Class-incremental learning (CIL) is a long-standing challenge in machine learning: models must keep learning new categories without erasing what they already know. The latest arXiv preprint tackles this by building on CLIP, a vision-language model whose multimodal pretraining yields features that transfer well across tasks.
The authors propose a method called hierarchical semantic tree anchoring, which uses the structure of class relationships to anchor CLIP's representations during incremental updates. While the abstract is brief, the approach appears to leverage semantic hierarchies to stabilize learning and reduce catastrophic forgetting.
The work sits at the intersection of continual learning and large pretrained models, suggesting that structured semantic knowledge can help adapt CLIP to evolving class sets.