Self-augmentation: Generalizing deep networks to unseen classes for few-shot learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) aims to classify new classes using limited data.
• Deep networks often memorize training data, hindering generalization in FSL.
• Existing FSL methods have not fully addressed deep network memorization issues.

Purpose of the Study:

• To propose a novel self-augmentation technique for few-shot learning.
• To improve the generalization ability of deep networks in few-shot scenarios.
• To address the problem of memorizing training statistics in deep networks.

Main Methods:

• Self-augmentation combining self-mix (regional dropout) and self-distillation.
• Self-mix replaces image patches to prevent learning dataset-specific structures.
• Auxiliary classifier branches enforce knowledge sharing and diverse learning.
• A local representation learner generates fake queries and novel weights for unseen classes.

Main Results:

• The proposed method significantly improves generalization ability in deep networks.
• Experimental results demonstrate superior performance on prevalent few-shot benchmarks.
• The approach effectively mitigates memorization of training statistics.

Conclusions:

• Self-augmentation offers a robust solution for few-shot learning challenges.
• The method enhances deep network performance by preventing overfitting to training data.
• This work advances the state-of-the-art in few-shot classification.