A spectral filtering approach to represent exemplars for visual few-shot classification.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot learning (FSL) commonly uses prototypes to represent category structures, acting as an inductive bias against overfitting.
• However, prototype-based methods struggle with categories lacking clear prototypes, potentially causing underfitting and limiting FSL performance.
• Exemplars offer an alternative representation for such challenging categories in FSL.

Purpose of the Study:

• To introduce Shrinkage Exemplar Networks (SENet), a novel approach for few-shot classification.
• To address the limitations of prototype-based FSL by incorporating exemplar-based representations.
• To enhance category representation by accounting for both the presence and absence of prototypes.

Main Methods:

• Propose Shrinkage Exemplar Networks (SENet) where category samples are adapted to shrink towards prototypes.
• Employ spectral filtering techniques to achieve sample shrinkage, enabling robust representation.
• Introduce a shrinkage exemplar loss function to replace traditional cross-entropy loss for improved sample information capture.

Main Results:

• Experimental validation on miniImageNet, tiered-ImageNet, and CIFAR-FS datasets.
• Demonstrated significant effectiveness of the proposed SENet method in few-shot classification tasks.
• Achieved superior performance compared to existing methods, particularly for categories with ambiguous prototypes.

Conclusions:

• SENet provides an effective approach for few-shot classification by adaptively representing categories.
• The shrinkage mechanism and novel loss function contribute to improved performance, especially in challenging FSL scenarios.
• The proposed method offers a promising direction for advancing few-shot learning research.