Few-Shot Meta Learners for Domain Adaptation

Abstract

Modern developments in Deep Learning and Machine Learning have shown great capacity to learn from labelled training assumption that training data and the data which the learning algorithm might see during testing or deployment have the same distribution. However this might not be true in most cases. We call this problem domainshift. In addition to the fact that it might not be possible to collect data from every possible domain gathering labelled data is expensive and resource consuming. So there is a need to build a learning algorithm that can adapt to new domains effectively and from small training samples. Hence, we propose a Meta Learning based approach using a Few Shot Model Agnostic Meta Learning(MAML) Algorithm to tackle problem of domain adaptation.