Multidomain Adaptation With Sample and Source Distillation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised multidomain adaptation leverages labeled source domains for unlabeled target tasks.
• Effective transfer learning relies on the quality, not just quantity, of training samples.
• Existing methods may not optimally select or weight source domain information.

Purpose of the Study:

• To propose a novel multidomain adaptation method named Sample and Source Distillation (SSD).
• To improve transfer learning performance by enhancing source sample quality and domain relevance.
• To address the challenge of efficiently utilizing knowledge from multiple labeled source domains for an unlabeled target domain.

Main Methods:

• Developed a two-step selective strategy for distilling source samples and ranking source domain importance.
• Constructed pseudo-labeled target domains and category classifiers to identify effective source samples.
• Employed a domain discriminator on selected source samples to estimate domain importance and merging weights.

Main Results:

• Successfully identified and distilled high-quality source samples and ranked source domains based on their relevance to the target task.
• Adapted multilevel distributions in a latent feature space using selected samples and ranked domains for effective knowledge transfer.
• Demonstrated superior performance of the proposed SSD method on real-world visual classification tasks.

Conclusions:

• The proposed Sample and Source Distillation (SSD) method effectively improves unsupervised multidomain adaptation.
• Selective sample distillation and source domain ranking are crucial for enhancing transfer learning performance.
• SSD offers a robust approach for leveraging knowledge from multiple labeled domains to unlabeled target tasks in visual classification.