Unsupervised Meta-Learning for Few-Shot Image Classification

Explore unsupervised meta-learning techniques for few-shot image classification in this 26-minute conference talk from NeurIPS 2019. Delve into the problem definition, task construction, and model-agnostic meta-learning approaches presented by researchers from the University of Central Florida. Learn about the UMTRA method, which utilizes random sampling and augmentation for task creation. Examine various augmentation techniques for datasets like Omniglot, ImageNet, and CelebA. Analyze ablation studies on Omniglot and Mini-ImageNet, and discover how video augmentation can be applied for self-supervision. Gain insights into feature visualization using t-SNE, particularly focusing on the last hidden layer. This presentation offers a comprehensive overview of cutting-edge approaches in unsupervised meta-learning for few-shot image classification tasks.

Intro
Outline
Problem definition
What is a task?
What we learn during meta-learning?
Model-agnostic meta-learning
Action recognition
Unsupervised Meta-learning with Tasks constructe by Random sampling and Augmentation UMTRA
Omniglot augmentation
Auto augmentation
ImageNet augmentation
Few-shot learning benchmarks
Ablation studies Omniglot
Ablation studies Mini-Imagenet
CelebA task generation
Video augmentation (self supervision)
Features visualization by t-SNE
t-SNE visualization of the last hidden layer