Feature extraction framework based on contrastive learning with adaptive positive and negative samples.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• High-dimensional data presents challenges in dimensionality.
• Contrastive learning is a prominent self-supervised technique for feature learning.
• Existing methods may lack robustness and adaptability in feature extraction.

Purpose of the Study:

• To propose a unified feature extraction framework using contrastive learning with adaptive samples (CL-FEFA).
• To enhance feature extraction for unsupervised, supervised, and semi-supervised learning tasks.
• To improve the compactness of intra-class samples and dispersion of inter-class samples.

Main Methods:

• Developed CL-FEFA framework utilizing adaptive positive and negative sample construction.
• Employed subspace sample structure information for dynamic sample generation, enhancing robustness to noise.
• Theoretically demonstrated maximization of mutual information for positive samples, capturing non-linear dependencies.

Main Results:

• CL-FEFA adaptively generates more accurate positive and negative samples for feature extraction.
• The framework extracts discriminative features, leading to more compact intra-class and dispersed inter-class embeddings.
• Numerical experiments confirmed CL-FEFA's superior performance over traditional and existing contrastive learning methods.

Conclusions:

• CL-FEFA offers a robust and effective approach to feature extraction across diverse learning paradigms.
• Adaptive sample selection and theoretical underpinnings provide significant advantages in handling high-dimensional data.
• The proposed method represents a notable advancement in unsupervised, supervised, and semi-supervised feature learning.