Joint Data Filtering and Labeling Using Gaussian Processes and Alternating Direction Method of Multipliers

Area of Science:

• Machine Learning
• Signal Processing
• Statistical Modeling

Background:

• Traditional sequence labeling methods often filter data independently before labeling, leading to suboptimal performance.
• Independent filtering can complicate and reduce the accuracy of subsequent labeling tasks in signal and image processing.
• Existing approaches face challenges in optimizing the interplay between filtering and labeling.

Purpose of the Study:

• To present a novel approach for sequence labeling that integrates filtering and labeling into a single, optimized process.
• To develop a method that jointly trains a Gaussian process classifier and estimates optimal filter coefficients.
• To improve the accuracy and efficiency of sequence labeling tasks.

Main Methods:

• A novel Bayesian modeling approach is employed, treating all unknowns as stochastic variables.
• Joint estimation of Gaussian process classifier coefficients and optimal filter coefficients.
• Utilizes variational inference for estimating unknowns and the Alternating Direction Method of Multipliers (ADMM) to link filtering and labeling.

Main Results:

• The proposed method demonstrates superior performance compared to existing approaches in both synthetic and real-world experiments.
• Simultaneous optimization of filtering and labeling leads to more accurate sequence labeling.
• The Bayesian framework provides a robust method for handling uncertainties in the data.

Conclusions:

• Jointly training a Gaussian process classifier and estimating filter coefficients offers a significant advancement in sequence labeling.
• The integration of Bayesian modeling and ADMM optimization provides an effective framework for simultaneous filtering and labeling.
• This approach enhances the overall performance and accuracy of sequence labeling applications.