Knot selection in sparse Gaussian processes with a variational objective function.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Sparse, knot-based Gaussian processes are effective scalable approximations for full Gaussian processes.
• Current methods for knot selection, while successful, can be computationally intensive due to simultaneous parameter optimization.

Purpose of the Study:

• To develop a more efficient method for selecting the number and locations of knots in sparse Gaussian processes.
• To address the lack of established methods for determining the optimal number of knots.

Main Methods:

• A one-at-a-time knot selection algorithm utilizing Bayesian optimization.
• Comparison against simultaneous knot optimization on benchmark datasets.

Main Results:

• The proposed Bayesian optimization method achieves competitive performance.
• The new method offers a significant reduction in computational cost compared to simultaneous optimization.

Conclusions:

• Bayesian optimization provides an efficient and effective approach for knot selection in sparse Gaussian processes.
• This method offers a practical alternative for optimizing sparse Gaussian process models.