Masked Embedding Modeling With Rapid Domain Adjustment for Few-Shot Image Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot classification faces challenges with limited labeled data and unknown distributions.
• Prototypical representation methods struggle with out-of-domain generalization, particularly with small support sets.

Purpose of the Study:

• To introduce a novel self-supervised generative technique to improve few-shot classification accuracy, especially for out-of-domain scenarios with small support sets.
• To develop a method for rapidly adapting few-shot learning models to new domains.

Main Methods:

• Proposed Masked Embedding Modeling for Few-Shot Learning (MEM-FS), a self-supervised generative technique using masked autoencoders to expand embedded support sets.
• Introduced Rapid Domain Adjustment (RDA), a self-supervised process for quick domain conditioning of MEM-FS.
• Applied MEM-FS+RDA to an inductive classifier backbone.

Main Results:

• MEM-FS+RDA significantly improved backbone performance on both out-of-domain and in-domain datasets.
• Achieved state-of-the-art performance on mini-imagenet, CVPR L2ID Classification Challenge, and IKEA-FS.
• Demonstrated the effectiveness of masked support embeddings for enhancing classification accuracy.

Conclusions:

• MEM-FS, augmented with RDA, offers a robust solution for few-shot classification challenges, particularly in out-of-domain settings.
• The proposed self-supervised generative approach enhances prototypical backbone models, leading to improved generalization.
• This work provides a significant advancement in few-shot learning, with practical implications for various classification tasks.