Experiments in cross-domain few-shot learning for image classification.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Cross-domain few-shot learning is crucial for practical applications where labeled data is scarce.
• Pretrained feature extractors offer a promising approach to address data limitations in new domains.

Purpose of the Study:

• To investigate optimal configurations of feature extractors and shallow classifiers for cross-domain few-shot learning.
• To evaluate the impact of feature extractor size, extraction stage, and normalization on performance.

Main Methods:

• Utilized ResNet-based feature extractors pretrained on ImageNet, applied to five diverse target domains.
• Experimented with logistic regression, linear discriminant analysis, and cosine similarity classifiers.
• Investigated feature vector normalization using various p-norms and incorporated multi-instance learning.

Main Results:

• Cosine similarity and logistic regression with L2-regularization achieved top classification performance.
• Linear discriminant analysis showed improved accuracy with L2-normalized features.
• Features from the penultimate stage of ResNet-101 and multi-instance learning yielded the highest accuracy across most domains.

Conclusions:

• The findings provide practical guidance for selecting feature extractors and classifiers in cross-domain few-shot learning.
• Optimal configurations involve specific ResNet stages, normalization techniques, and classifier choices for enhanced performance.