Discriminative Transfer Learning Using Similarities and Dissimilarities.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning (TL) addresses the challenge of building effective classification models for categories with scarce training data.
• It leverages knowledge from related source categories that possess abundant training data.

Purpose of the Study:

• To propose a novel discriminative transfer learning (DTL) method.
• To enhance classification performance by integrating knowledge from both target and source categories.

Main Methods:

• The DTL method utilizes sparse reconstruction residuals as a core discriminant.
• It enhances discrimination by comparing residuals from positive and negative dictionaries, incorporating both correlated and anti-correlated source categories.
• A Wilcoxon-Mann-Whitney statistic-based cost function is employed for selecting dictionaries with imbalanced data.
• Two parallel boosting processes are applied to positive and negative data distributions.

Main Results:

• The proposed DTL method consistently outperformed existing state-of-the-art TL methods on two image classification databases.
• The method demonstrated efficient runtime performance.

Conclusions:

• The developed DTL method offers a robust and efficient approach for transfer learning.
• It effectively leverages source category knowledge to improve classification accuracy, especially in low-data scenarios.