Gaussian Scale Mixture

Gaussian Scale Mixtures (GSMs) model data as a mixture of Gaussian distributions with varying variances, offering a flexible framework for representing heavy-tailed and non-Gaussian data. Current research focuses on improving parameter estimation efficiency, particularly for over-specified models, using algorithms like Expectation-Maximization (EM) and novel approaches such as Exponential Location Updates (ELU). Applications span diverse fields, including image quality assessment, blind source separation (e.g., speech enhancement), and robust deep learning through network pruning, demonstrating GSM's versatility in handling complex data structures and improving model performance. The development of efficient algorithms and the extension of GSMs to various model architectures highlight its growing importance in signal processing and machine learning.