Causal multi-label learning for image classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-label image classification presents challenges due to diverse supervision signals.
• Existing methods for multi-label learning often involve complex procedures and lack intuitive interpretations.
• Previous approaches focused on label-related image areas or label co-occurrence.

Purpose of the Study:

• To propose a novel approach for causal image classification with multi-label learning.
• To overcome limitations of existing methods by incorporating causal inference.
• To develop an elegant and effective method with low computational cost.

Main Methods:

• Introduced causal multi-label learning (CMLL) incorporating causal inference.
• Employed a multi-class attention module for selecting multiple objects from images.
• Applied causal intervention to learn causal relationships between labels.

Main Results:

• Demonstrated significant improvement in prediction performance.
• Showcased low computational cost and few parameters required for the approach.
• Verified effectiveness through extensive tests and ablation studies.

Conclusions:

• CMLL offers an effective and computationally efficient solution for causal multi-label image classification.
• The proposed method enhances prediction accuracy without substantially increasing training or inference times.
• Causal inference provides a beneficial framework for addressing challenges in multi-label learning.