From Instance to Metric Calibration: A Unified Framework for Open-World Few-Shot Learning

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Robust few-shot learning (RFSL) addresses noisy labels but often assumes noise is in-domain.
• Real-world scenarios frequently involve out-of-domain noise, creating a gap for existing RFSL methods.
• Open-world few-shot learning (OFSL) addresses the simultaneous presence of in-domain and out-of-domain noise.

Purpose of the Study:

• To propose a unified framework for comprehensive calibration in open-world few-shot learning.
• To effectively mitigate the impact of both in-domain and out-of-domain noise in few-shot datasets.
• To enhance the robustness and performance of models in complex, noisy few-shot learning settings.

Main Methods:

• A dual-networks structure combining a contrastive network and a meta network.
• Instance-wise calibration using a novel prototype modification strategy with reweighting.
• Metric-wise calibration employing a novel metric fusing spatial metrics from both networks.

Main Results:

• The proposed framework effectively mitigates noise from both feature and label spaces.
• Demonstrated robustness and superiority across various open-world few-shot learning settings.
• The method successfully handles simultaneous in-domain and out-of-domain noise.

Conclusions:

• The developed unified framework provides a robust solution for open-world few-shot learning.
• The instance and metric calibration strategies effectively address complex noise scenarios.
• The approach significantly advances the state-of-the-art in handling noisy few-shot data.