Partial Domain Adaptation Without Domain Alignment.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) transfers knowledge between domains with identical label spaces.
• Partial Domain Adaptation (PDA) addresses scenarios where the source domain's label space includes the target domain's.
• Existing UDA methods struggle with PDA due to label space discrepancies, leading to negative transfer.

Purpose of the Study:

• To propose a novel approach for Partial Domain Adaptation (PDA) that avoids complex domain alignment.
• To theoretically and empirically validate a new method for knowledge transfer in scenarios with non-identical label spaces.

Main Methods:

• Theoretical analysis bounding target risk by model smoothness and domain discrepancy.
• Instantiating model smoothness as Intra-Domain Structure Preserving (IDSP).
• Evaluating IDSP on benchmark datasets without relying on domain alignment.

Main Results:

• IDSP demonstrates superior performance compared to state-of-the-art PDA methods (e.g., ~+10% on Cl → Rw, ~+8% on Ar → Rw).
• The proposed method achieves significant improvements without employing domain alignment.
• IDSP is shown to be complementary to domain alignment techniques in standard UDA.

Conclusions:

• Intra-Domain Structure Preserving (IDSP) offers an effective and simpler alternative for Partial Domain Adaptation (PDA).
• The findings suggest that focusing on intra-domain structure is crucial for successful knowledge transfer in PDA.
• IDSP presents a promising direction for advancing domain adaptation research, particularly in practical, non-identical label space scenarios.