Missing Modality Transfer Learning via Latent Low-Rank Constraint.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Transfer learning typically requires target domain data during training, which is often unavailable for multi-modal datasets.
• Cross-modal transfer learning faces challenges when modalities differ between source and target domains or when target data is absent.

Purpose of the Study:

• To develop a unified framework for knowledge transfer across databases and modalities, even with missing target modality data.
• To enhance recognition performance in multi-modal transfer learning scenarios where target data is inaccessible during training.

Main Methods:

• Introduced a latent factor to model the underlying structure of missing modalities.
• Employed bidirectional transfer learning for data alignment across modalities and databases.
• Developed an efficient latent low-rank transfer learning algorithm with theoretical guarantees.

Main Results:

• The proposed method successfully inherited knowledge from auxiliary databases and source modalities.
• Demonstrated significant improvements in recognition performance despite the absence of target modality data during training.
• Experimental results validated the effectiveness of the approach in multi-modal knowledge transfer with missing target modalities.

Conclusions:

• The developed latent low-rank transfer learning framework effectively addresses the challenge of missing target modalities in cross-modal transfer learning.
• The method offers a robust solution for leveraging available data to improve recognition accuracy in complex multi-modal scenarios.
• This work advances the field of transfer learning by enabling knowledge transfer even when target domain information is incomplete.