Feature Mixture on Pre-Trained Model for Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) requires robust feature extraction for novel object recognition with limited data.
• Training effective feature extractors (backbones) is challenging due to time constraints and bias towards base category features.
• Existing backbones struggle to generalize to novel categories, focusing on textures rather than essential representations.

Purpose of the Study:

• To propose a novel feature mixture (FM) operation to improve FSL performance without retraining backbones.
• To enhance the generalizability and diversity of training samples by mixing pre-trained features.
• To address the limitations of traditional backbone training in FSL.

Main Methods:

• A feature mixture operation is applied to pre-trained, fixed features.
• Part of the feature map values from novel categories are replaced with content from other feature maps.
• Feature similarities are used to constrain the mixture operation, guiding the classifier to focus on relevant novel object representations.

Main Results:

• The feature mixture operation significantly improves FSL accuracy across five benchmark datasets in both inductive and transductive settings.
• On Mini-ImageNet, FM achieved 3.8% and 4.2% accuracy gains with 1 and 5 training samples, respectively, compared to the baseline.
• The proposed method enhances other FSL approaches that rely on backbone training.

Conclusions:

• The feature mixture operation offers an effective and computationally efficient solution for improving FSL.
• This approach enhances feature generalizability and helps overcome the limitations of pre-trained backbones.
• FM is a versatile technique that can be integrated with existing FSL methods to boost their performance.