Weight Copy and Low-Rank Adaptation for Few-Shot Distillation of Vision Transformers

arXiv:2404.09326v1 Announce Type: cross
Abstract: Few-shot knowledge distillation recently emerged as a viable approach to harness the knowledge of large-scale pre-trained models, using limited data and computational resources. In this paper, we propose a novel few-shot feature distillation approach for vision transformers. Our approach is based on two key steps. Leveraging the fact that vision transformers have a consistent depth-wise structure, we first copy the weights from intermittent layers of existing pre-trained vision transformers (teachers) into shallower architectures (students), where …

abstract arxiv computational copy cs.ai cs.cv cs.lg data distillation feature few-shot harness key knowledge low low-rank adaptation novel paper pre-trained models resources scale transformers type vision vision transformers