Model Recycling Framework for Multi-Source Data-Free Supervised Transfer Learning
Researchers have developed a novel framework for transfer learning that operates without access to the original source data. This "model recycling" approach allows for parameter-efficient training by identifying and reusing subsets of related pre-trained models. The framework is designed to function in both white-box and black-box scenarios, enabling Model as a Service providers to create libraries of efficient models for multi-source data-free supervised transfer learning. AI
IMPACT Enables more efficient model training and deployment by reducing reliance on original datasets.