Reliable Few-Shot Learning Under Dual Noises.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) adapts pre-trained models using limited data.
• Existing FSL methods struggle with in-distribution (ID) and out-of-distribution (OOD) noise in support and query samples.
• Noise amplification during adaptation leads to unreliable predictions in FSL.

Purpose of the Study:

• To propose DEnoised Task Adaptation (DETA++) for robust few-shot learning.
• To enhance the reliability of FSL models in the presence of dual noise (ID and OOD).
• To improve task adaptation and prediction accuracy with limited labeled data.

Main Methods:

• Developed a Contrastive Relevance Aggregation (CoRA) module for sample weighting.
• Introduced a clean prototype loss and noise entropy maximization loss for robust adaptation.
• Employed a memory bank and Local Nearest Centroid Classifier (LocalNCC) for noise-robust predictions.
• Utilized an Intra-class Region Swapping (IntraSwap) strategy to rectify class prototypes.

Main Results:

• DETA++ demonstrates significant improvements in noise-robust task adaptation.
• The proposed methods effectively mitigate the adverse effects of dual noises in FSL.
• Experiments confirm the effectiveness and flexibility of the DETA++ framework.

Conclusions:

• DETA++ offers a reliable solution for few-shot learning challenges posed by noisy data.
• The method enhances model adaptability and predictive performance in open-world settings.
• The approach provides a flexible and effective framework for noise-robust FSL.