Sample-Centric Feature Generation for Semi-Supervised Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised few-shot learning leverages limited labeled data and abundant unlabeled data to boost model generalization.
• Existing methods often use label propagation or pseudo-labeling, which can lead to distribution gaps between pseudo-labels and real data.

Purpose of the Study:

• To address the coarse-grained feature distribution issue in pseudo-labeled data for semi-supervised few-shot image classification.
• To propose a novel sample-centric feature generation (SFG) approach to improve model performance.

Main Methods:

• The proposed SFG approach generates derivative features around pseudo-labeled samples to enrich intra-class diversity.
• It employs a semi-supervised meta-generator and sample-centric constraints for compact and discriminative features.
• A reliability assessment (RA) metric is introduced to mitigate the impact of outliers.

Main Results:

• The SFG approach effectively enriches intra-class feature diversity while maintaining inter-class discriminability.
• The reliability assessment metric successfully reduces the influence of outlier generated features.
• Experiments demonstrate significant improvements on challenging one- and few-shot image classification benchmarks.

Conclusions:

• The sample-centric feature generation approach offers a robust solution for semi-supervised few-shot image classification.
• This method effectively bridges the distribution gap between pseudo-labeled and real query data.
• The proposed technique enhances model generalization by better utilizing unlabeled data.