How to Trust Unlabeled Data? Instance Credibility Inference for Few-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep learning models achieve high performance but require extensive labeled data.
• Few-shot learning (FSL) addresses challenges with limited labeled data in real-world scenarios.
• Existing FSL methods often rely on meta-learning or data augmentation.

Purpose of the Study:

• To introduce a novel statistical approach, Instance Credibility Inference (ICI), for few-shot visual recognition.
• To leverage unlabeled data to enhance the performance of models trained with limited labeled examples.
• To provide a theoretically grounded method for selecting credible pseudo-labeled instances.

Main Methods:

• Repurposing self-taught learning to predict pseudo-labels for unlabeled instances.
• Employing a (Generalized) Linear Model (LM/GLM) with incidental parameters to assess instance credibility.
• Ranking pseudo-labeled instances based on the sparsity of incidental parameters along a regularization path.
• Iteratively augmenting the training set with high-credibility instances.

Main Results:

• The ICI approach effectively utilizes unlabeled data for few-shot visual recognition.
• Demonstrated significant improvements on benchmark datasets like miniImageNet, tieredImageNet, CIFAR-FS, and CUB.
• Theoretical guarantees for collecting correctly predicted pseudo-labeled instances under specific conditions.
• Outperforms existing methods in few-shot learning settings.

Conclusions:

• Instance Credibility Inference (ICI) offers a robust statistical alternative for few-shot learning.
• The method successfully exploits unlabeled data, reducing reliance on extensive manual annotation.
• ICI enhances the scalability of deep learning models to real-world long-tail categories.
• The approach is validated by extensive experiments and theoretical analysis.