The meta-learning method for the ensemble model based on situational meta-task.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Meta-learning is crucial for few-shot learning, enabling generalization to new tasks.
• Current meta-learning methods lack consideration for the relationship between meta-tasks and novel tasks, limiting meta-knowledge utility.
• Variations in initial models lead to inconsistent performance in few-shot learning scenarios.

Purpose of the Study:

• To propose a meta-optimization method addressing limitations in existing meta-learning approaches.
• To enhance meta-knowledge transfer for improved generalization in few-shot learning.
• To leverage situational meta-task construction and multi-model cooperation for better performance on novel tasks.

Main Methods:

• Developed a situational meta-task construction method during meta-training to select more effective task sets.
• Implemented a meta-optimization ensemble model approach during meta-testing for cooperative learning.
• Minimized inter-model prediction loss to facilitate effective collaboration among multiple models.

Main Results:

• The proposed method was applied to few-shot character and image recognition datasets.
• Experimental results demonstrated the effectiveness of the approach in few-shot classification tasks.
• The method achieved good performance, indicating improved generalization capabilities.

Conclusions:

• The proposed meta-optimization strategy effectively improves few-shot learning performance.
• Situational meta-task construction and model cooperation are key to enhancing meta-knowledge.
• Future work aims to extend the method for scenarios with completely unseen novel tasks.