SIN: Semantic Inference Network for Few-Shot Streaming Label Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Streaming label learning requires extensive new data for emerging labels in multilabel classification.
• Real-world scenarios often provide only limited annotated data for new labels in dynamic environments.

Purpose of the Study:

• To introduce and investigate few-shot streaming label learning (FSLL) for modeling new labels with minimal data.
• To develop a meta-learning framework that leverages past label knowledge for efficient adaptation to new labels.

Main Methods:

• Propose the Semantic Inference Network (SIN), a meta-learning framework for FSLL.
• Utilize label semantic representation to regularize the output space and acquire labelwise meta-knowledge via gradient-based meta-learning.
• Incorporate a novel label decision module with a meta-threshold loss for optimal confidence threshold determination.

Main Results:

• SIN effectively adapts to new FSLL tasks using only a few annotated examples.
• The semantic inference mechanism theoretically constrains the hypothesis space, reducing overfitting and improving generalizability.
• Empirical results demonstrate SIN's superior performance compared to existing state-of-the-art methods on FSLL tasks.

Conclusions:

• FSLL is a viable approach for handling emerging labels with limited data in multilabel classification.
• The proposed SIN framework offers an effective solution for few-shot adaptation in streaming label learning.
• SIN's ability to infer semantic correlations and optimize decision thresholds enhances performance and generalizability.