Combining Model-Agnostic Meta-Learning and Transfer Learning for Regression.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Model adaptation is crucial for applying machine learning to new tasks.
• Existing methods like model-agnostic meta-learning (MAML) and transfer learning have limitations.
• The influence of data distribution shifts on adaptation performance is not fully understood.

Purpose of the Study:

• To investigate the impact of data distribution differences on the performance of machine learning adaptation methods in regression.
• To develop robust ensemble schemes that mitigate the negative effects of distribution shifts.
• To enhance the stability and generalizability of machine learning models across diverse tasks.

Main Methods:

• Conducted experiments on regression problems to analyze the effect of varying data distributions between source and target tasks.
• Developed and evaluated ensemble schemes that combine multiple adaptation strategies.
• Utilized sinusoidal fitting, virtual reality motion prediction, and temperature forecasting as benchmark regression tasks.

Main Results:

• Data distribution differences between old and new tasks significantly influence the relative performance of different adaptation methods.
• The proposed ensemble schemes demonstrated superior and more stable performance compared to individual methods across various distribution shifts.
• The ensemble approach achieved the best results in most of the evaluated regression problems.

Conclusions:

• The degree of data distribution difference is a critical factor affecting the success of machine learning model adaptation.
• Ensemble methods provide a robust solution for handling diverse data distribution shifts in regression tasks.
• The developed ensemble schemes offer improved and reliable performance for a wide range of real-world applications.