Uncertainty Estimation With Neural Processes for Meta-Continual Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Evaluating uncertainties in dynamic data streams is critical, especially during events like pandemics.
• Neural Process Families (NPFs) integrate Gaussian Processes (GPs) and Neural Networks (NNs) for uncertainty prediction, suitable for limited data scenarios.
• Existing NPF models lack robust solutions for continual learning with strict data access constraints.

Purpose of the Study:

• To introduce a novel framework, member meta-continual learning with neural process (MCLNP), for enhanced uncertainty estimation in evolving data streams.
• To enable dual-level uncertainty estimation: local point uncertainties and global function evolution uncertainties (p(z)).
• To address catastrophic forgetting in continual learning by incorporating a coreset memory buffer.

Main Methods:

• Developed MCLNP by combining NPFs with meta-continual learning principles.
• Implemented a coreset mechanism to retain previous knowledge and mitigate forgetting.
• Investigated the relationship between global uncertainties, intratask diversity, and model complexity.

Main Results:

• Successfully estimated prediction uncertainties across various evolving data types (abrupt, gradual, recurrent shifts).
• Demonstrated effectiveness on 1-D, 2-D datasets, and a novel spatial-temporal COVID dataset.
• Achieved superior performance over baseline methods in terms of likelihood and rapid adaptation to data stream evolution.

Conclusions:

• MCLNP provides a robust solution for uncertainty estimation in dynamic and evolving environments.
• The coreset strategy effectively supports continual learning by preventing knowledge loss.
• The proposed method shows significant promise for applications requiring adaptive predictions in complex, changing data landscapes.