Target-Directed Progressive Gradient Adjusting for transfer learning

Yijun Bei, Kewei Gao, Zhuoyang Zhao, Sen Lin, Erteng Liu, Zunlei Feng

Abstract: Highlights•Proposed Target-directed Progressive Gradient Adjusting (TPGA) to enhance transfer learning by progressively adjusting source gradients with target guidance, ensuring optimal convergence.•Theoretically proved TPGA converges to the ideal optimum in the target domain.•Extensive experiments on image classification tasks showed TPGA outperforms existing methods, achieving SOTA results.