Learning Calibrated Class Centers for Few-Shot Classification by Pair-Wise Similarity

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Metric-based few-shot classification methods learn from limited data by clustering support samples.
• Existing approaches suffer from inaccurate class center approximations due to small support sets, leading to biased inference.

Purpose of the Study:

• To address the issue of inaccurate class center approximation in metric-based few-shot learning.
• To enhance the performance of few-shot image classification by reducing inference bias.

Main Methods:

• Propose a Pair-wise Similarity Module (PSM) for class center calibration.
• PSM captures semantic correlations between support and query samples.
• PSM enhances discriminative regions in support representations.

Main Results:

• The PSM significantly improves the performance of conventional metric-based few-shot classification models.
• Experiments conducted on four benchmark few-shot image classification datasets demonstrate effectiveness.
• The PSM is a plug-and-play module compatible with existing methods.

Conclusions:

• Class center calibration using the PSM effectively reduces approximation errors in few-shot learning.
• The proposed PSM offers a simple yet powerful enhancement for metric-based few-shot classification.
• This work advances few-shot image classification accuracy through improved representation learning.