FACT: A Simple and Efficient Framework for Active Finetuning
Researchers have introduced FACT, a novel framework designed to enhance the efficiency and effectiveness of active finetuning for pretrained models. This approach addresses the issue of feature distortion during finetuning by employing a three-phase hierarchical strategy. Experiments show FACT achieves significant performance gains, particularly on smaller datasets, with improvements exceeding 20% on ViT models for several benchmarks. AI
IMPACT Enhances model adaptation efficiency, particularly beneficial for scenarios with limited labeled data.