Adaptive Sparse Gaussian Process.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Adaptive learning is crucial for nonstationary environments, requiring machines to forget outdated data distributions.
• Efficient algorithms need compact, computationally inexpensive model updates for online parameter tuning.
• Current solutions inadequately address these adaptive learning challenges.

Purpose of the Study:

• To propose the first adaptive sparse Gaussian process (GP) addressing computational efficiency and nonstationarity.
• To develop a method for compact model updates with minimal computational cost.
• To enable effective online parameter updating in dynamic environments.

Main Methods:

• Reformulated a variational sparse GP (VSGP) algorithm incorporating a forgetting factor for adaptivity.
• Developed a novel approach to update only a single inducing point and model parameters per new data sample.
• Focused on simplifying model inference for efficient online processing.

Main Results:

• The proposed algorithm demonstrates fast convergence in its inference process.
• Achieved efficient model updates with a single inference iteration, even in highly nonstationary settings.
• Showcased strong performance in predictive posterior mean and confidence interval estimation.

Conclusions:

• The adaptive sparse GP effectively handles nonstationary environments by forgetting past data.
• The method offers computational efficiency through compact, single-iteration model updates.
• Outperforms state-of-the-art approaches in modeling predictive posteriors and confidence intervals.