Semantic-Aware Pseudo-Labeling for Unsupervised Meta-Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised meta-learning utilizes clustering-based pseudo-labeling for model-agnostic learning from unlabeled data.
• Existing methods face challenges with clustering noise and semantic chaos in pseudo-labels, limiting meta-learning effectiveness.

Purpose of the Study:

• To optimize the pseudo-labeling process in unsupervised meta-learning to generate semantic-like pseudo-labels.
• To bridge the performance gap between unsupervised and supervised meta-learning.

Main Methods:

• Minimizing the inter-to-intra-class similarity ratio to create clustering-friendly embedding features during encoding.
• Proposing a semantic-stability index and a Semantic-aware Pseudo-label Reassignment mechanism for improved pseudo-label quality during clustering.

Main Results:

• The proposed encoding method generates clustering-friendly embeddings, reducing clustering noise.
• The semantic-aware reassignment mechanism effectively generates semantic-like pseudo-labels.
• Integration with MAML and EP algorithms showed significant improvements on few-shot benchmarks.

Conclusions:

• The developed approach is model-agnostic and enhances unsupervised meta-learning by producing high-quality pseudo-labels.
• The method achieves state-of-the-art results and even surpasses supervised methods in certain few-shot learning tasks.