.png){kind=icon}
.png){kind=icon}
Return
Target-Directed Progressive Gradient Adjusting for transfer learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.111731.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning has revolutionized various tasks and achieved impressive performance, relying on abundant data. Transfer learning has emerged as a promising approach to address the challenges posed by limited data in target domains. Existing techniques typically involve transferring knowledge from pre-trained models in the source domain to the target domain through regularization strategies. These strategies constrain the fine-tuning process within a restricted optimization space, often with limited supervision from target domain samples. While these regularization techniques contribute to good performance, they can also result in the retention of significant source domain knowledge, hindering the optimization toward an ideal solution for the target domain. This paper presents a novel technique, Target-directed Progressive Gradient Adjusting (TPGA), for transfer learning. Unlike the traditional pre-training and fine-tuning paradigm, TPGA progressively adjusts the gradient direction of source domain samples during each iteration, guided by a small number of target domain samples. A theoretical proof of the convergence of TPGA is provided from a gradient optimization perspective. A new transfer learning framework based on TPGA is then proposed. Extensive experiments on image classification and segmentation tasks demonstrate that TPGA outperforms state-of-the-art methods, achieving superior transfer learning performance, especially in scenarios with limited target domain data.
Keywords:
Image classification
Image segmentation
Transfer learning