Concept-consistent continual learning stops saliency explanation drift
Original title: Semantic Concept Steering Breaks the Explanation Drift Loop in Continual Learning
C3L anchors regularization to offline class concepts, and authors report gains on five benchmarks against over 20 baselines.
The paper says pixel-level saliency regularization in continual learning forms a self-referential loop that reinforces spurious attention. C^3L instead regularizes toward class-consistent semantic concepts discovered offline and independent of model state. Approximate LRP relevance conservation makes concept regions crowd out background correlations without an explicit penalty. Concept discovery and spatial grounding are automated.
Unconfirmed: State-of-the-art claims are the authors' evaluation against the baselines they selected.
- Accuracy can stay stable while attention shifts to spurious features.
- Example concepts include curved beak and striped pattern.
- No human-annotated concept sets are required.
- Reported metrics cover forgetting, explanation quality, and out-of-distribution robustness.
From the original article
1 min read at huggingface.co
Catastrophic forgetting in continual learning (CL) manifests not only as accuracy degradation but also as explanation drift: models maintain predictive accuracy while silently shifting attention from diagnostic features to spurious correlations.
By Yehonatan Elisha, Oren Barkan, Noam Koenigstein · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.