Diverse expected gradient active learning for relative attributes.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semantic understanding of images and videos is crucial for human-machine communication.
• Manual labeling of attributes in large datasets is labor-intensive and time-consuming.
• Existing active learning methods often select samples serially, limiting efficiency for multi-attribute learning.

Purpose of the Study:

• To propose a novel batch-mode active learning method for efficient multi-attribute learning.
• To address the limitations of serial sample selection in active learning.
• To improve the process of discovering and labeling informative data samples.

Main Methods:

• Introduced a batch-mode active learning method named diverse expected gradient active learning.
• Integrated informativeness analysis (expected pairwise gradient length) and diversity analysis (diverse gradient angle).
• Employed a two-step procedure for optimizing informativeness and diversity, and a heuristic for imbalanced distributions.

Main Results:

• Demonstrated the effectiveness and efficiency of the proposed approach through empirical evaluations.
• The method successfully forms a diverse batch of informative queries.
• The heuristic method effectively suppresses imbalanced multiclass distributions.

Conclusions:

• The proposed diverse expected gradient active learning method enhances efficiency in learning multiple attributes.
• Batch-mode active learning is more efficient than serial mode for multi-attribute learning.
• The approach offers a promising solution for semantic understanding in large-scale image and video datasets.