Deep and Structured Robust Information Theoretic Learning for Image Analysis.

Area of Science:

• Machine Learning
• Information Theory
• Computer Vision

Background:

• Data representation tasks often suffer from uncertainties like missing or noisy labels.
• Existing methods may not effectively handle these uncertainties in discriminative learning.
• Maximizing mutual information is a key objective for robust data representation.

Purpose of the Study:

• To propose a robust information theoretic (RIT) model for reducing label uncertainties in data representation.
• To simultaneously learn data transformation and discriminative classification functions.
• To enhance the performance of discriminative tasks by maximizing mutual information in the latent space.

Main Methods:

• Developed a general RIT model framework.
• Implemented three RIT variants: linear subspace embedding, deep transformation, and structured sparse learning.
• Applied RIT and deep RIT to image categorization and structured sparse RIT to brain MRI segmentation.

Main Results:

• The RIT and deep RIT models demonstrated effective performance in image categorization tasks on benchmark datasets.
• Structured sparse RIT showed promise in medical image analysis for brain tissue segmentation.
• The model successfully reduced uncertainties associated with missing and noisy labels.

Conclusions:

• The proposed RIT model offers a robust approach to handle label uncertainties in discriminative learning.
• RIT models are versatile and applicable to various tasks, including image categorization and medical image segmentation.
• Future work can explore further applications and extensions of the RIT framework.