On Adaptive Learning Framework for Deep Weighted Sparse Autoencoder: A Multiobjective Evolutionary Algorithm.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep weighted sparse autoencoders (AEs) require effective hyperparameter optimization for performance.
• Existing methods may lack efficiency in adapting sparsity constraints for deep neural networks.

Purpose of the Study:

• To establish an adaptive learning framework for deep weighted sparse AEs.
• To enhance hyperparameter optimization using a multiobjective evolutionary algorithm (MOEA).

Main Methods:

• Developed an adaptive learning framework integrating a deep weighted sparse AE with MOEA.
• Introduced weighted sparsity for flexible constraint design.
• Implemented a divide-and-conquer strategy and a sharing scheme to improve MOEA efficiency.

Main Results:

• The adaptive learning framework demonstrates effectiveness in optimizing deep sparse AEs.
• Experimental results validate the framework's generality across various sparse models, convolutional AEs, and the VGG-16 network.
• The framework shows applicability in blind image quality assessment tasks.

Conclusions:

• The proposed adaptive learning framework offers an effective approach for deep sparse AE optimization.
• The integration of MOEA with specific strategies enhances learning efficiency and precision.
• The framework's successful application highlights its practical utility in image analysis and other domains.