An α -Divergence-Based Approach for Robust Dictionary Learning.

Area of Science:

• Signal Processing
• Machine Learning
• Computer Vision

Background:

• Dictionary learning (DL) is crucial for signal processing and machine learning.
• Traditional DL often uses quadratic loss, which is sensitive to outliers.
• Robustness in DL is needed for handling non-Gaussian noise and improving stability.

Purpose of the Study:

• To introduce a robust sequential dictionary learning algorithm.
• To improve upon existing dictionary learning methods by incorporating a novel robust loss function.
• To ensure inference stability even with significant deviations from standard noise models.

Main Methods:

• Developed a dictionary learning algorithm based on maximum likelihood estimation.
• Incorporated a robust loss function derived from α-divergence, replacing the standard quadratic loss.
• Utilized a sequence of penalized rank-1 matrix approximation problems with l1-norm regularization.
• Employed a block coordinate descent approach for parameter estimation.

Main Results:

• The proposed robust loss function belongs to the class of redescending M-estimators.
• The algorithm demonstrates superior performance compared to other robust DL methods.
• Efficacy validated on tasks including digit recognition, background removal, and grayscale image denoising.

Conclusions:

• The presented robust sequential DL algorithm offers improved stability and performance.
• The use of α-divergence-based robust loss is effective for handling noise.
• The algorithm shows promise for various image processing and machine learning applications.