What question did this study set out to answer?

This research investigates how integrating visual and semantic data affects few-shot image classification.

April 19, 2026SymmetryOpen Access

Asymmetric Cross-Modal Prototypical Networks for Few-Shot Image Classification

View Full Paper

Ask AI

Bookmark

Authors

An experimental study reveals the benefits of integrating visual and semantic information in few-shot learning.

This research investigates how integrating visual and semantic data affects few-shot image classification.
Developed Asymmetric Cross-Modal Prototypical Networks (ACM-ProtoNet) using frozen CLIP text encoders.
Implemented a scalar fusion gate to balance visual and semantic features.
Conducted experiments on miniImageNet with one-shot and five-shot classifications, applying both ablation studies and statistical testing.
Found a significant improvement in five-shot learning (+2.12 percentage points) with cross-modal integration.
Noted that random text features reduced performance by 7.4–8.1 percentage points.
Observed that text-only prototypes performed near-random (≈20% accuracy).
Ablation studies confirmed semantic binding was essential, showing consistent degradation when class names were shuffled.

Xie et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark