SLTRN: Sample-level transformer-based relation network for few-shot classification.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot classification aims to identify novel categories using minimal labeled data.
• Traditional Relation Networks (RN) for few-shot classification do not fully leverage contextual information within the support set, limiting comparison accuracy.
• There is a need for improved methods that can effectively mine relationships among support samples for better classification.

Purpose of the Study:

• To address the limitations of existing Relation Networks in few-shot classification.
• To introduce a novel approach, the Sample-level Transformer-based Relation Network (SLTRN), for enhanced few-shot classification.
• To improve the ability to recognize novel categories by effectively utilizing support set contextual information.

Main Methods:

• Reformulated the learning of relationships between query and support samples as a sequence-to-sequence (seq2seq) problem.
• Developed a Sample-level Transformer-based Relation Network (SLTRN) incorporating sample-level self-attention.
• Employed self-attention mechanisms to mine potential relationships among support classes, enhancing comparison capabilities.

Main Results:

• SLTRN achieved performance comparable to state-of-the-art methods on benchmark datasets.
• Demonstrated particular strength in the 1-shot setting, achieving 52.11% accuracy on miniImageNet and 67.55% accuracy on CUB.
• Ablation experiments confirmed the effectiveness and identified optimal settings for SLTRN.

Conclusions:

• The proposed Sample-level Transformer-based Relation Network (SLTRN) effectively enhances few-shot classification.
• Leveraging sample-level self-attention significantly improves the mining of relationships within the support set.
• SLTRN offers a promising advancement for recognizing novel categories with limited labeled data.