An Efficient Sparse Bayesian Learning Algorithm Based on Gaussian-Scale Mixtures

Area of Science:

• Machine Learning
• Signal Processing
• Computational Statistics

Background:

• Sparse Bayesian Learning (SBL) offers strong generalization but suffers from high computational cost due to matrix inversions.
• Large-scale datasets exacerbate the computational bottleneck in traditional SBL algorithms.
• Efficient SBL methods are crucial for practical applications in signal and image processing.

Purpose of the Study:

• To develop an efficient Sparse Bayesian Learning (SBL) algorithm with reduced computational complexity.
• To address the limitations of traditional SBL for large-scale data applications.
• To enhance both computational efficiency and model sparsity in SBL.

Main Methods:

• A Gaussian-scale mixture prior model is employed for efficient SBL.
• A surrogate function approximates posterior density, avoiding matrix inversions.
• A joint cost function and block coordinate descent within a majorization-minimization framework solve the optimization problem.

Main Results:

• The proposed SBL algorithm achieves O(n^2) computational complexity per iteration.
• Experimental results demonstrate superior performance in sparse signal recovery and image reconstruction.
• The approach shows significant improvements in computational time and estimation error compared to existing methods.

Conclusions:

• The developed efficient SBL algorithm effectively overcomes the computational limitations of traditional SBL.
• The method provides a balance between computational efficiency and model sparsity.
• This approach is highly effective for large-scale sparse signal recovery and image reconstruction problems.