A multi-fidelity tabular prior-data fitted network model for accurate prediction and uncertainty quantification.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Accurate prediction from feature-label datasets is crucial for drug discovery, diagnostics, and climate science.
• Challenges include limited data, high dimensionality, and multi-fidelity scenarios.

Purpose of the Study:

• To develop a general-purpose multi-fidelity model for enhanced prediction accuracy and uncertainty quantification.
• To integrate low- and high-fidelity data effectively for improved machine learning performance.

Main Methods:

• Developed Multi-Fidelity Tabular Prior-Data Fitted Network (MFTabPFN) using a hierarchical transformer architecture.
• Integrated low- and high-fidelity data to capture cross-fidelity correlations.
• Implemented an active learning framework for scalable model refinement.

Main Results:

• MFTabPFN demonstrated superior performance compared to state-of-the-art methods across diverse tasks.
• Achieved significant improvements in prediction accuracy and uncertainty quantification.
• Showcased versatility in handling both single- and multi-fidelity datasets.

Conclusions:

• MFTabPFN offers robust prediction and uncertainty quantification capabilities.
• The model is a promising tool for data-driven discovery in various scientific applications.
• Active learning enhances scalability for resource-intensive tasks.