A novel meta-learning framework: Multi-features adaptive aggregation method with information enhancer.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels at image classification but requires substantial training data.
• Few-shot classification tasks present challenges in effective feature representation and classifier learning with limited samples.

Purpose of the Study:

• To propose a multi-features adaptive aggregation meta-learning (MFAML) method for few-shot image classification.
• To enhance feature extraction and classifier performance in low-data scenarios.

Main Methods:

• Developed MFAML with three modules: feature extraction, information enhancer, and multi-features adaptive aggregation classifier (MFAAC).
• Utilized deconvolutional layers in the information enhancer to improve sample utilization and feature extraction.
• Integrated features from multiple convolutional layers via MFAAC for adaptive label prediction.
• Employed a hybrid loss and model-agnostic meta-learner (MAML) optimization strategy.

Main Results:

• The information enhancer and MFAAC modules, connected by a hybrid loss, yield superior feature representation.
• Experimental results on benchmark datasets demonstrate MFAML's superiority over other few-shot classification methods.
• The framework shows effective generalization performance improvement.

Conclusions:

• MFAML effectively addresses the challenges of few-shot image classification by improving feature representation and classifier learning.
• The proposed method offers a promising approach for scenarios with limited training data.
• MFAML demonstrates significant advantages over existing representative methods in few-shot learning tasks.