Latent Elastic-Net Transfer Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Transfer learning methods often seek a common subspace to reduce domain discrepancy, but this doesn't guarantee optimal classification.
• Existing approaches may fail to find the most discriminative subspace for accurate predictions.

Purpose of the Study:

• To propose a novel transfer learning method, Latent Elastic-Net Transfer Learning (LET), for enhanced domain adaptation.
• To simultaneously learn both a latent subspace and a discriminative subspace for improved classification performance.

Main Methods:

• LET minimizes Maximum Mean Discrepancy (MMD) to interlace data in a latent subspace, decoupling inputs and outputs for compact representation.
• A low-rank constraint based matrix elastic-net regression is employed to learn a discriminative subspace capturing intra-class correlations.

Main Results:

• The proposed LET method effectively learns a discriminative subspace for classification.
• Experiments on visual domain adaptation tasks demonstrate the superiority of the LET method over existing approaches.

Conclusions:

• LET offers a more effective approach to domain adaptation by learning both latent and discriminative subspaces.
• The method ensures better discriminative alignment, leading to superior classification performance in visual domain adaptation.