Deep Unsupervised Active Learning via Matrix Sketching.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Existing unsupervised active learning methods often struggle with complex, non-linear data structures due to reliance on linear models and data reconstruction loss.
• This limitation hinders their effectiveness in real-world applications with intricate data patterns.

Purpose of the Study:

• To propose a novel deep unsupervised active learning method, ALMS (Active Learning with Matrix Sketching), designed to handle complex non-linear data structures.
• To improve sample selection for manual labeling in unsupervised settings, enhancing classification task performance.

Main Methods:

• ALMS employs a deep auto-encoder for data embedding into a latent space.
• Matrix Sketching is utilized to summarize embedded data characteristics with a compact sketch.
• Representative samples are selected to approximate this sketch, preserving key information while reducing parameters.

Main Results:

• ALMS demonstrates superior performance on both single-label and multi-label classification tasks compared to state-of-the-art methods.
• The method effectively alleviates model overfitting and efficiently handles large datasets.
• An auxiliary self-supervised task (real/fake sample classification) further enhances encoder representation ability.

Conclusions:

• ALMS offers a robust solution for unsupervised active learning, particularly for data with complex non-linear structures.
• The Matrix Sketching approach provides a self-supervised signal, improving model learning and representation.
• ALMS represents a significant advancement in efficient and effective active learning for classification.