Transfer Learning with Uncertainty Quantification: Random Effect Calibration of Source to Target (RECaST).

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Transfer learning adapts models trained on one dataset for use with another.
• Current transfer learning methods often lack uncertainty quantification.
• Fine-tuning pre-trained neural networks is a common but limited approach.

Purpose of the Study:

• To develop a statistical framework for transfer learning with uncertainty quantification.
• To introduce the RECaST (Recalibration with Cauchy Random Effects for Statistical Transfer) framework.
• To demonstrate the validity and robustness of RECaST across different model types.

Main Methods:

• Developed a statistical framework, RECaST, utilizing a Cauchy random effect for model recalibration.
• Mathematically and empirically validated RECaST for linear models, ensuring prediction set coverage.
• Numerically illustrated robustness for nonlinear models, showing resilience to asymptotic approximations.

Main Results:

• RECaST provides uncertainty quantification for transfer learning predictions.
• The framework is source-model agnostic and does not require source data access.
• Demonstrated RECaST's effectiveness through simulations and real-world hospital data analysis.

Conclusions:

• RECaST offers a statistically sound and versatile approach to transfer learning.
• The inclusion of uncertainty quantification addresses a critical gap in existing methods.
• RECaST shows promise for reliable model adaptation across diverse populations and data types.