Multi-Stage Meta-Learning for Few-Shot with Lie Group Network Constraint.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models often face overfitting when training data is limited.
• Meta-learning offers a solution by leveraging knowledge from similar tasks to adapt to new ones with few samples.
• Existing meta-learning methods using Shallow Neural Networks (SNNs) and Euclidean gradient descent can be inefficient and inaccurate.

Purpose of the Study:

• To propose a novel meta-learning model, Multi-Stage Meta-Learning (MSML), to overcome limitations in adapting to new tasks with scarce data.
• To enhance the feature extraction and parameter update processes in meta-learning.
• To improve the accuracy and efficiency of meta-learning models in few-shot learning scenarios.

Main Methods:

• Developed the Multi-Stage Meta-Learning (MSML) model.
• Constrained neural network parameters to the Stiefel manifold to ensure stable gradient descent.
• Implemented a novel approach to accelerate the adaptation process in meta-learning.

Main Results:

• The proposed MSML model demonstrated improved accuracy on the mini-ImageNet dataset.
• Achieved better performance under 5-way 1-shot and 5-way 5-shot learning conditions compared to existing methods.
• Showcased more stable and efficient gradient descent for meta-learners.

Conclusions:

• MSML effectively addresses the overfitting problem in deep learning with limited labeled samples.
• Constraining networks to the Stiefel manifold enhances meta-learner stability and accelerates adaptation.
• The proposed method represents a significant advancement in few-shot learning and meta-learning research.