Guided Random Projection Boosts Continual Representation Learning on Pretrained Models
arXiv 2603.19145·low signal
This paper shows that adding a guided random projection layer to pretrained model-based continual learning significantly improves performance by reducing representational interference between tasks. The approach is lightweight and compatible with existing pretrained backbones without requiring additional finetuning infrastructure. It addresses the catastrophic forgetting problem for practitioners deploying models that need to acquire new tasks without retraining from scratch.