Subdomain Adaptation With Manifolds Discrepancy Alignment.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Transfer learning aims to reduce domain divergence, but existing methods often overlook local subdomain differences.
• Domains can comprise multiple subdomains with distinct divergence patterns, necessitating a localized approach.

Purpose of the Study:

• To propose a novel transfer learning framework that accounts for local subdomain divergence.
• To introduce a method for aligning data distributions within local manifold representations across domains.

Main Methods:

• Representing subdomains using low-dimensional manifolds.
• Developing manifold maximum mean discrepancy (M3D) to measure local distribution discrepancies.
• Introducing transfer with manifolds discrepancy alignment (TMDA) to couple manifold discovery with M3D minimization.

Main Results:

• TMDA framework successfully instantiated for subspace learning (linear and nonlinear) and deep learning.
• Experimental results demonstrate the efficacy of TMDA across various transfer learning tasks.

Conclusions:

• TMDA offers a promising approach for addressing local domain divergence in transfer learning.
• Accounting for subdomain discrepancies via manifold alignment enhances transfer learning performance.