DILC-ESAE: Data-Info envelope stacked autoencoder on correlation among samples rather than themselves.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Stacked autoencoders (SAEs) are effective for structured classification.
• Existing SAEs overlook inter-sample correlation, limiting classification performance.
• Inter-sample correlation can significantly improve feature separability and classification accuracy.

Purpose of the Study:

• To propose a novel SAE that effectively mines correlations between samples.
• To introduce the Data-Info envelope stacked autoencoder (DILC-ESAE) for enhanced classification.
• To improve classification accuracy by modeling inter-sample correlations.

Main Methods:

• The proposed DILC-ESAE integrates a Data-Info layer construction model (DILC) to extract local and global sample correlations, creating envelope samples.
• An embedded stacked autoencoder (ESAE) fuses original and correlation-based features during training and within the network architecture.
• The method focuses on deep feature extraction from sample correlations, not just individual samples.

Main Results:

• The DILC-ESAE demonstrated superior performance compared to existing representative SAEs in experiments.
• The approach successfully extracts hierarchical deep features by modeling correlations among samples.
• The DILC component shows potential for application in other autoencoders and deep neural networks.

Conclusions:

• The DILC-ESAE offers a novel approach to deep feature extraction by utilizing inter-sample correlations.
• This method enhances classification accuracy by capturing relationships between data points.
• The DILC-ESAE framework provides a valuable reference for future deep learning architectures.