Incremental Ensemble Gaussian Processes.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Gaussian processes (GPs) excel at nonlinear function learning and uncertainty quantification.
• Standard GP methods often use a single kernel, limiting adaptability for sequential, time-critical data.
• Online kernel adaptation is crucial for dynamic environments.

Purpose of the Study:

• To develop an incremental ensemble (IE-) GP framework for online kernel adaptation.
• To enable GPs to handle time-varying functions and improve performance in adversarial settings.
• To explore online unsupervised learning for dimensionality reduction using IE-GP.

Main Methods:

• An ensemble of GP learners, each with a unique kernel, is assembled using data-adaptive weights.
• Random feature-based approximation enables scalable online prediction and model updates.
• Structured dynamics are modeled within the assembler and learners for time-varying functions.
• Regret analysis is used to evaluate performance in adversarial conditions.

Main Results:

• The IE-GP framework demonstrates effective online kernel adaptation for sequential data.
• The dynamic variant of IE-GP successfully models time-varying functions.
• Performance analysis via regret confirms robustness in adversarial settings.
• Online dimensionality reduction is achieved effectively under the IE-GP framework.

Conclusions:

• The proposed IE-GP framework offers a scalable and adaptive solution for sequential data analysis.
• IE-GP enhances uncertainty quantification and prediction accuracy in dynamic and adversarial environments.
• The framework shows promise for applications requiring real-time learning and dimensionality reduction.