Learning representations from unlabeled images for few-shot learning

Paper Title:

Unsupervised Representation Learning to Aid Semi-Supervised Meta Learning

Published on:

19 October 2023

Primary Category:

Machine Learning

Paper Authors:

Atik Faysal,

Mohammad Rostami,

Huaxia Wang,

Avimanyu Sahoo,

Ryan Antle

Bullets

Key Details

•

Proposes unsupervised representation learning for few-shot classification

•

Uses data augmentation to create episodes from unlabeled images

•

Transfers learned representations to supervised few-shot learning

•

Avoids need for full manual labeling of training data

•

Enables effective semi-supervised few-shot learning

Explore the topics in this paper

few-shot learning

image classification

representation learning

semi-supervised learning

unsupervised learning

AI generated summary

Learning representations from unlabeled images for few-shot learning

This paper proposes an unsupervised representation learning method to improve few-shot image classification. It uses data augmentation to generate training episodes from unlabeled images. The learned representations are transferred to a supervised few-shot learner for fast adaptation and improved accuracy. Key benefits are avoiding extensive manual labeling and enabling semi-supervised learning.