Metric Learning: Learning Distances in Feature Space

Contents

• 1. Introduction
• 2. What is Metric Learning
• 3. Applications of Metric Learning

1. Introduction

Metric Learning is a machine learning technique that aims to learn a distance function so that samples from the same class are closer together, while samples from different classes are farther apart in the feature space.

Illustration of metric learning: distinguishing distances in embedding space

This technique plays an important role in tasks such as recognition, classification, and similarity search.

2. What is Metric Learning

Metric Learning learns a distance function or an embedding space where the distance between points with the same label is smaller than the distance between points with different labels.

Popular methods include:

• Contrastive loss
• Triplet loss
• N-pair loss

3. Applications of Metric Learning

• Face recognition
• Image retrieval
• Few-shot learning
• Speaker verification
• Radio frequency fingerprint classification. Part 1 where I train RiftNet is here, and Part 2 about the openset problem will be updated later.
• And many other applications…

Example of metric learning application in signal recognition