CS5720 - Week 8
Slide 153 of 160

Few-Shot Learning

What is Few-Shot Learning?

Few-Shot Learning is the ability to learn new concepts from just a few examples, similar to how humans can quickly recognize new objects after seeing them only once or a few times.
Key Terms:

N-way K-shot: N classes, K examples per class
Support Set: Few labeled examples for learning
Query Set: Test examples to classify
Meta-Learning: Learning how to learn efficiently
🧠 Think of it like...
A child who can recognize a new animal after seeing just one picture in a book. They use their prior knowledge about animals to quickly understand this new concept!

Few-Shot Learning Methods

  • 📏
    Metric Learning
    Learn a similarity metric to compare examples
  • 🔄
    Model-Agnostic (MAML)
    Learn initialization that adapts quickly to new tasks
  • 🎯
    Prototypical Networks
    Create class prototypes and classify by proximity
  • 🧠
    Memory Networks
    Store and retrieve relevant past experiences

5-Way 1-Shot Learning Example

Support Set
1 example per class (5 classes)
🐱
🐶
🐦
🐠
🐸
One example each: Cat, Dog, Bird, Fish, Frog
Query Example
New example to classify
🐕
Which class does this belong to?
Prediction
Model's classification
🐶
Predicted: Dog (95% confidence)
Challenge: Learn to classify new examples using only one example per class
Prepared by Dr. Gorkem Kar