INTERLACE: A New Framework for Efficient Pruning and Adaptive Fine-Tuning of Vision-Language Models

Introduction / Main Floor: INTERLACE: A New Framework for Efficient Pruning and Adaptive Fine-Tuning of Vision-Language Models

The research team from UC Santa Barbara proposes the INTERLACE framework. Through triplet layer importance analysis and interleaved pruning strategy, it achieves that after pruning 25% of the layers of large vision-language models, the performance remains at 88.9% while using only 1% of the training data, bringing breakthrough progress to the field of model compression.