Targeted Transfer Learning: Leveraging Optimal Transport for Enhanced Knowledge Transfer

Machine learning models are often built based on the notion that the labeled training data follows the same underlying distribution of the testing dataset. Often this assumption does not hold and the trained model's performance becomes limited during the deployment in the target environment. To address this problem, we present a transfer learning-based targeted transfer learning (TTL) approach that employs the optimal transport distance (OTD) metric to facilitate knowledge transfer between the source and target domains. TTL method uses a target-side adaptation methodology by intentionally modifying the source domain data using a small number of target domain samples to improve transfer performance between source and target datasets. Experimental results on the benchmark datasets show the effectiveness of the proposed approach without re-training the source model on the target environment, which contributes to the selection of an appropriate source model and improves the performance of the target model.