Semi-supervised learning for training CNNs with Few Data

Abstract

Although Deep Learning has successfully been applied to many fields, it relies on large amounts of data. In this work we focus on two different research lines within the context of image classification that try to deal with this problem. a) The first part of the project is focused on Active Learning (AL), which is an extensive field within Machine Learning that tries to reduce the amount of labeling work by inter- actively querying the most informative samples from a large dataset. Most of the AL literature is based on uncertainty sampling methods which do not perform so well when applied to neural networks. In this project we present a density estimation approach for Active Learning that overcomes some of the sampling limitations re- lated to the uncertainty-based methods. b) The second part of the project is focused on a very recent field within deep learning called one-shot learning, which aims to correctly classify samples by just seeing one or few training samples from each class. In this work we present a simple non-linear learnable metric for one-shot learning that overcomes most of the state of the art results obtained with simple methods and is competitive in terms of accuracy to more complex ones. We also present a meta-learner architecture based on Graph Neural Networks for one-shot learning.