Hybrid Fine-Tuning Strategy for Few-Shot Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot classification (FSC) aims to train models with limited labeled data.
• Current FSC methods often struggle with optimal fine-tuning, leading to accuracy or generalization issues.
• Existing approaches inadequately address the trade-off between task-specific adaptation and generalization.

Purpose of the Study:

• To propose a novel hybrid fine-tuning strategy (HFT) for few-shot classification.
• To enhance feature extraction and label prediction capabilities in low-data regimes.
• To improve both accuracy on target tasks and generalization to unseen samples.

Main Methods:

• Introduced a hybrid fine-tuning strategy (HFT) comprising two modules: Few-Shot Linear Discriminant Analysis (FSLDA) and Adaptive Fine-Tuning (AFT).
• FSLDA initializes classifier parameters by constructing an optimal linear classification function for few-shot conditions.
• AFT employs an adaptive termination rule to determine optimal training epochs, preventing overfitting and enhancing generalization.

Main Results:

• The proposed HFT method demonstrated consistent performance improvements across various few-shot classification benchmarks.
• Experiments on mini-ImageNet and tiered-ImageNet validated the effectiveness of FSLDA and AFT modules.
• The strategy achieved superior results compared to existing fine-tuning methods under diverse sample sizes and layer freezing policies.

Conclusions:

• The hybrid fine-tuning strategy (HFT) effectively addresses limitations in current few-shot classification fine-tuning.
• FSLDA and AFT provide a robust framework for optimizing model adaptation and generalization in low-data scenarios.
• This approach offers a significant advancement for few-shot learning tasks requiring high accuracy and robust generalization.