Dual-Representation-Based Autoencoder for Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Natural Language Processing

Background:

• Domain adaptation addresses learning in unlabeled target domains using labeled source domains.
• Existing autoencoder methods for domain adaptation often damage class-discriminative information while aligning distributions.
• This can lead to decreased performance as samples from different classes become closer.

Purpose of the Study:

• To propose a novel Dual-Representation Autoencoder (DRAE) for domain adaptation.
• To overcome the limitation of damaged class-discriminative information in existing autoencoder-based approaches.
• To learn dual-domain-invariant representations for improved domain adaptation performance.

Main Methods:

• DRAE learns global representations to maximize inter-class distance and minimize distribution discrepancies.
• DRAE extracts local representations to preserve class-discriminative information within each class.
• Dual representations are constructed by aligning global and local representations with weighted contributions.

Main Results:

• Extensive experiments were conducted on three text and two image datasets.
• DRAE was compared against 12 state-of-the-art domain adaptation methods.
• The proposed DRAE demonstrated significant effectiveness in domain adaptation tasks.

Conclusions:

• The Dual-Representation Autoencoder (DRAE) effectively addresses the issue of lost class-discriminative information in domain adaptation.
• DRAE successfully learns domain-invariant representations while preserving crucial class-specific details.
• The method shows superior performance across diverse text and image datasets, validating its efficacy.