Lecture Slides

05_zero-shot_learning.pdf

Few Shot Learning Motivation

What if we lack a large dataset for training a neural network?

very limited amount of data for certain objects

Examples include medical imaging, translations for rare languages, ecology, robotics and personalized education.

Training Set: Dataset used to train the base model with classes that have sufficient labeled examples

Support Set: Small set of labeled examples for new classes seen during testing

Query: New examples that need to be classified into one of the new classes using the support set

• First, you train your model on the large training set (dogs, cats, birds) to learn general features about animals
• Then, you use the support set (few examples of zebras and giraffes) to quickly adapt your model to recognize these new classes by comparing query set to the support set using some sorts of similarity measure.
• The class of the query set should be one of the classes in the support set.
• defined by N-way (number of classes) and k-shot (number of samples in each class)

<aside> 🧠

The support set should contain examples from classes that weren't in the training set, but these new classes should belong to the same general domain or distribution as the training data.

• Shared underlying features - The training and support set should share fundamental features that transfer across classes. For example, if you trained on various animals, the support set might contain new animal species (not seen during training) rather than vehicles or furniture.
• Similar data characteristics - The image quality, lighting conditions, backgrounds, etc., should be similar between training and support sets to avoid domain shift problems.
• Conceptual relatedness - The new classes in the support set should be conceptually related to the training classes in a way that allows for meaningful knowledge transfer. </aside>

Datasets

Metric Learning vs. Meta Learning

Also called nonparametric and parametric methods

Nonparametric (Non-Meta)

Do not rely on a fixed set of parameters. Instead, use the support set directly at inference time.

Parametric methods (Meta)

Use only a fixed number of parameters learned during training.

Siamese Network