Leveraging Bilateral Correlations for Multi-Label Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multi-label few-shot learning (ML-FSL) aims to classify images with multiple labels using limited training data.
• Existing ML-FSL methods model instance-label correlations but assume uniform importance, limiting knowledge extraction from few examples.
• This uniform importance assumption acts as a bottleneck, hindering the full utilization of correlations in ML-FSL.

Purpose of the Study:

• To propose a unified framework, Bilateral Correlation Reconstruction (BCR), to address the limitations of uniform importance assumptions in ML-FSL.
• To enable networks to mine underlying instance and label correlations with varying importance from both instance-to-label and label-to-instance perspectives.

Main Methods:

• BCR refines category prototypes by reweighting images based on their instance-importance degree, calculated from instance-category similarity (instance-to-label perspective).
• BCR smooths image labels by recovering latent label-importance, considering the integrated topology of all samples within a task (label-to-instance perspective).

Main Results:

• Experimental results on multiple benchmarks demonstrate the effectiveness of the proposed BCR framework.
• BCR significantly outperforms existing ML-FSL methods, indicating improved performance in few-shot multi-label image classification.

Conclusions:

• The proposed Bilateral Correlation Reconstruction (BCR) framework effectively addresses the bottleneck of uniform importance assumptions in ML-FSL.
• BCR enhances the mining of instance and label correlations by considering varying importance, leading to superior performance in multi-label few-shot learning tasks.