Class‑Incremental Learning (CIL) enables a model to acquire new classes over time while preserving previously learned knowledge. However, many real‑world scenarios are compositional: visual concepts arise from combinations of primitives such as states (e.g., red, shiny) and objects (e.g., shoe, car). Conventional CIL treats each composition as an independent class, ignoring shared primitives and offering no mechanism to keep primitive representations stable as new compositions appear. Consequently, parameter updates entangle primitives, causing representation drift and degrading performance on earlier compositions.
Compositional Incremental Learning (composition‑IL) formalizes this setting by introducing new state‑object compositions across tasks. To address it, we propose Layer‑Specific Adaptation with Prototype‑Anchored Adapter‑Space Transport (LSA‑PAST), a parameter‑efficient solution.
LSA integrates Low‑Rank Adaptation (LoRA) into a frozen Vision Transformer (ViT) and equips depth‑specialized adapters for states, objects, and their compositions. This design allows fine‑grained adjustments at different layers without altering the backbone weights.
PAST maps past, anchored prototypes into the current adapter space via an affine transform fitted on the current‑task data. The process is rehearsal‑free, ensuring that old prototypes remain compatible with newly adapted adapters and thus mitigating drift.
Extensive experiments on several composition‑IL benchmarks show that LSA‑PAST achieves a superior accuracy‑forgetting trade‑off, setting a new state‑of‑the‑art performance.
Review