Online sparse Gaussian process regression and its applications.

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Modeling

Background:

• Gaussian processes (GPs) are powerful probabilistic models but can be computationally intensive.
• Existing GP inference methods often struggle with large datasets and real-time applications.
• Efficiently updating GP models with new data is crucial for online learning.

Purpose of the Study:

• To develop a novel Gaussian process inference algorithm for efficient online learning.
• To apply the new algorithm to real-world problems like head pose estimation and visual tracking.
• To demonstrate the algorithm's efficiency, accuracy, and generalization capabilities.

Main Methods:

• Introduced the online sparse matrix Gaussian processes (OSMGP) algorithm, leveraging sparse Gram matrices with local support kernels.
• Utilized Givens rotations for efficient maintenance and updating of the sparse Cholesky factor.
• Implemented matrix downdates for constant-time operations and online hyperparameter estimation.

Main Results:

• Achieved linear time complexity for updates, scaling efficiently with dataset size.
• Demonstrated constant-time operations using matrix downdates, enabling efficient data discarding.
• Showcased robust and accurate performance in head pose estimation and visual tracking tasks.
• Validated the algorithm's generalization ability through online learning experiments.

Conclusions:

• The OSMGP algorithm offers a significant advancement in efficient Gaussian process inference.
• Its online nature and computational efficiency make it suitable for real-time computer vision applications.
• The method provides a scalable and accurate solution for dynamic learning tasks.