Learning to Discover Knowledge: A Weakly-Supervised Partial Domain Adaptation Approach.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation leverages source domain knowledge but collecting high-quality source data is challenging.
• Real-world data often contains noisy labels, necessitating methods for generalized domain adaptation.
• Weakly-supervised partial domain adaptation (WS-PDA) addresses transferring classifiers from noisy source domains to unlabeled target domains.

Purpose of the Study:

• To propose a novel domain adaptation approach, self-paced transfer classifier learning (SP-TCL), for WS-PDA.
• To address the challenges of discovering knowledge from noisy and unlabeled data and adapting it across domains.
• To establish a strong baseline for generalized domain adaptation tasks.

Main Methods:

• SP-TCL employs a self-paced learning scheme to identify a suitable classifier for the target domain.
• A prudent loss function is designed to extract faithful knowledge from noisy source and unlabeled target domains.
• Knowledge adaptation is achieved by iteratively excluding source examples during training in a self-paced manner.

Main Results:

• Extensive evaluations on benchmark datasets demonstrate the efficacy of SP-TCL.
• SP-TCL significantly outperforms state-of-the-art approaches in generalized domain adaptation tasks.
• The proposed method offers a simple yet effective solution for WS-PDA.

Conclusions:

• SP-TCL provides an effective solution for weakly-supervised partial domain adaptation by leveraging noisy source data.
• The self-paced learning framework enables robust knowledge discovery and cross-domain adaptation.
• This approach serves as a strong baseline for various generalized domain adaptation challenges.