LSDT: Latent Sparse Domain Transfer Learning for Visual Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation is crucial for applying models to new data distributions.
• Heterogeneous data presents challenges due to cross-domain distribution mismatch.
• Existing methods struggle with complex nonlinear shifts and noise.

Purpose of the Study:

• To propose a novel reconstruction-based transfer learning method for domain adaptation.
• To enhance visual categorization of heterogeneous data.
• To address cross-domain distribution mismatch and nonlinear subspace shifts.

Main Methods:

• Latent Sparse Domain Transfer (LSDT) using ℓ1-norm sparse coding for target domain data reconstruction.
• Joint learning model for simultaneous optimization of sparse coding and subspace representation.
• Kernel-based generalization for nonlinear subspace shifts in reproduced kernel Hilbert space.

Main Results:

• LSDT jointly learns latent space and reconstruction for optimal subspace transfer.
• Robust domain adaptation achieved by removing noise/outliers via sparse subspace clustering.
• Effective handling of highly nonlinear domain shifts, such as varying face poses.

Conclusions:

• The proposed LSDT method significantly outperforms state-of-the-art representation-based domain adaptation techniques.
• The kernel-based extension effectively tackles nonlinear domain shifts.
• LSDT offers a robust and versatile solution for heterogeneous data adaptation and visual categorization.