Structure Guided Deep Neural Network for Unsupervised Active Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Unsupervised active learning aims to select representative samples for labeling in unsupervised settings.
• Existing methods often use linear models, which struggle with real-world nonlinear data and can ignore cluster structures, especially in imbalanced datasets.

Purpose of the Study:

• To propose a novel deep unsupervised active learning framework that addresses the limitations of existing linear models.
• To improve sample selection by considering nonlinearity, global structure, cluster structure, and local structure.

Main Methods:

• A deep neural network is used to learn a nonlinear embedding of the data into a latent space.
• A selection block utilizes a self-supervised learning strategy to choose representative samples in the latent space.
• The framework preserves global data structure, captures cluster structure for imbalanced data, and maintains local structure for precise embedding.

Main Results:

• The proposed framework effectively selects representative samples by learning nonlinear embeddings.
• It successfully handles imbalanced data by capturing cluster structures.
• Preservation of global and local data structures leads to improved model performance.

Conclusions:

• The novel deep unsupervised active learning framework demonstrates superior performance compared to state-of-the-art methods.
• The method's ability to model nonlinearity and preserve data structures makes it effective for real-world applications.