Label smoothing and task-adaptive loss function based on prototype network for few-shot learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Prototype networks face challenges with unreliable label information.
• Loss function hyperparameters often fail to adapt to dynamic image feature changes.

Purpose of the Study:

• To develop a method combining label smoothing and adaptive hyperparameters for improved few-shot image classification.
• To enhance the robustness and accuracy of prototype networks in few-shot learning scenarios.

Main Methods:

• Image label information is regularized using label smoothing.
• Image feature distance matrices are fused with loss function hyperparameters using logarithmic operations.
• Hyperparameters are dynamically associated with smoothed labels and distance matrices for classification.

Main Results:

• The proposed method demonstrated a 2%-3% improvement in classification accuracy on miniImageNet, FC100, and tieredImageNet datasets.
• Compared to methods with unsmoothed labels and fixed hyperparameters, the flexible approach yielded better results.
• The method effectively suppresses interference from false labels and enhances classification accuracy.

Conclusions:

• Combining label smoothing and flexible hyperparameters significantly improves few-shot learning classification accuracy.
• The proposed method offers a robust solution for handling noisy labels and adapting to feature variations in image classification.