Negatives Make a Positive: An Embarrassingly Simple Approach to Semi-Supervised Few-Shot Learning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Semi-Supervised Few-Shot Learning (SSFSL) addresses training classifiers with minimal labeled data and available unlabeled data.
• Existing SSFSL methods often struggle with the inherent data scarcity in few-shot scenarios, limiting adaptability to new tasks.

Purpose of the Study:

• To develop a simple yet effective approach for generating accurate pseudo-labels in SSFSL.
• To augment limited support sets using predicted negative and positive pseudo-labels to enhance classifier performance.

Main Methods:

• Proposed an indirect learning strategy to predict accurate negative pseudo-labels for unlabeled data.
• Iteratively refined pseudo-labels by excluding negative ones to derive a positive pseudo-label for each sample.
• Integrated both negative and positive pseudo-labels to augment the limited support set.

Main Results:

• The proposed method achieved significant accuracy improvements on four benchmark datasets for SSFSL.
• Demonstrated superior performance compared to existing state-of-the-art SSFSL methods.
• The approach requires minimal code and uses off-the-shelf operations.

Conclusions:

• The developed pseudo-labeling technique effectively complements limited support sets in SSFSL.
• The method shows strong adaptability and generalization, functioning as a plug-and-play module for existing SSFSL techniques and linear models.