Training multi-source domain adaptation network by mutual information estimation and minimization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Multi-Source Domain Adaptation (MSDA) trains models on multiple labeled datasets for unlabeled target data.
• A key challenge is adapting to differing joint distributions across datasets.
• Existing methods struggle with effective distribution alignment.

Purpose of the Study:

• To propose a novel approach for Multi-Source Domain Adaptation (MSDA).
• To align source and target joint distributions in the latent feature space.
• To improve the performance of neural networks on unlabeled target data.

Main Methods:

• Estimating and minimizing mutual information within the network's latent feature space.
• Formulating mutual information estimation as a convex optimization problem for a guaranteed global optimum.
• Developing a novel algorithm for domain adaptation.

Main Results:

• The proposed algorithm statistically outperforms existing methods on public datasets.
• Demonstrated effective alignment of source and target joint distributions.
• Achieved superior classification performance on unlabeled target data.

Conclusions:

• Minimizing mutual information is an effective strategy for MSDA.
• The proposed convex optimization approach provides a globally optimal solution.
• The method offers a significant advancement in cross-domain learning for neural networks.