Information set supported deep learning architectures for improving noisy image classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models excel in supervised learning but are prone to overfitting.
• Uncertainty from data biases, class imbalance, or noise degrades deep learning performance.
• A need exists for robust deep learning architectures resilient to various data imperfections.

Purpose of the Study:

• To introduce Information-Set Deep Learning (ISDL) architectures.
• To address the limitations of standard deep learning models in handling uncertainty.
• To develop robust deep learning solutions for real-world data challenges.

Main Methods:

• Developed four variants of Information-Set Deep Learning (ISDL) architectures.
• Integrated information set theory with deep learning principles.
• Described ISDL architectures, learning algorithms, and analytic workflows.

Main Results:

• ISDL models demonstrated efficient handling of noise-dominated uncertainty.
• Evaluated ISDL models against standard architectures using a noise-corrupted dataset.
• ISDL models significantly outperformed peer architectures in performance.

Conclusions:

• ISDL architectures offer a robust solution for deep learning.
• The proposed models effectively mitigate performance degradation caused by data uncertainty.
• ISDL represents a significant advancement in developing resilient deep learning systems.