Unsupervised Few-Shot Feature Learning via Self-Supervised Training.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Few-shot learning (FSL) is crucial but current methods predominantly use supervised learning, requiring extensive labeled data.
• Unsupervised learning offers a more natural approach, mirroring cognitive processes and showing promise in various ML tasks.

Purpose of the Study:

• To propose a novel unsupervised feature learning method for few-shot learning.
• To develop a model that reduces dependency on large labeled datasets for effective FSL.

Main Methods:

• The proposed model employs two alternating processes: progressive clustering and episodic training.
• Progressive clustering generates pseudo-labeled data for creating episodic tasks.
• Episodic training utilizes these tasks to refine feature representations and train the few-shot learner.

Main Results:

• The model demonstrated strong generalization capabilities across diverse downstream few-shot learning tasks on benchmark datasets like Omniglot and MiniImageNet.
• Achieved effective performance in few-shot person re-identification on a newly created dataset, FS-Market1501.

Conclusions:

• The proposed unsupervised method effectively addresses limitations of supervised few-shot learning.
• This approach shows significant potential for real-world applications, including person re-identification, by improving feature learning from limited data.