RESEARCH · RESEARCH · #2313
LSA-PAST: parameter-efficient adapters with prototype transport for compositional incremental learning
The paper (arXiv:2610.27203v1) introduces LSA-PAST, a method that integrates LoRA-style, depth-specialized adapters into a frozen ViT for states, objects and compositions, and uses a rehearsal-free affine transform (PAST) to map past prototypes into the current adapter space so old prototypes remain compatible as new compositions arrive; experiments report state-of-the-art accuracy–forgetting trade-offs on compositional incremental learning benchmarks.
KEY POINTS
- The paper (arXiv:2610.27203v1) introduces LSA-PAST, a method that integrates LoRA-style, depth-specialized adapters into a frozen ViT for states, objects and compositions, and uses a rehearsal-free affine transform (PAST) to map past prototypes into the current adapter space so old prototypes remain compatible as new compositions arrive; experiments report state-of-the-art accuracy–forgetting trade-offs on compositional incremental learning benchmarks.
- It targets representation drift in compositional incremental learning by stabilizing primitive (state/object) representations and enabling rehearsal-free retention of prior compositions.
- LSA-PAST: Layer-Specific Adaptation and Prototype Transport for Compositional Incremental Learning
WHY IT MATTERS
It targets representation drift in compositional incremental learning by stabilizing primitive (state/object) representations and enabling rehearsal-free retention of prior compositions.