Few Shot Learning with Code - Meta Learning - Prototypical Networks

This video addresses one of the biggest drawbacks of classical deep learning, the requirement for a large amount of data. Few-Shot Learning has become a popular method for many researchers to deal with this issue in recent years. The goal of few-shot learning (FSL) is to teach a machine to do something new with a significantly small amount of data. I mentioned multiple popular algorithms that are used to solve FSL using the Meta-Learning framework. Then explained one of the popular algorithms called Prototypical Networks (NIPS’17, 3000 citations) along with how to train and test FSL on Omniglot dataset in Google Colab using PyTorch. Prototypical Networks classifies new classes (not part of training data/classes) based on their similarity to a small number of examples per class. This is one of the simplest algorithms to solve FSL. There are some other little more advanced algorithms like MAML, Meta-LSTM, and Reptile. I’ll cover them in future videos. Stay tuned. Thanks! -------------------- ✅👍📸 Subscribe to the Channel 👉 -------------------- Code Colab: Paper: -------------------- Chapters 0:00 Intro: Issues with Classical Deep Learning 0:51 Alternative to Classical Learning: Few-Shot learning 2:18 Classical Learning vs Meta-Learning 5:47 Concept: Prototypical Networks 7:00 Code: Prototypical Networks - Train & Test 13:34 End -------------------- #fewshotlearning #metalearning #prototypicalnetworks #computervision