Meta-Learning with Domain Adaptation for Few-Shot Learning under Domain Shift

Doyen Sahoo, Hung Le, Chenghao Liu, Steven C. H. Hoi

Abstract: Few-Shot Learning (learning with limited labeled data) aims to overcome the limitations of traditional machine learning approaches which require thousands of labeled examples to train an effective model. Considered as a hallmark of human intelligence, the community has recently witnessed several contributions on this topic, in particular through meta-learning, where a model learns how to learn an effective model for few-shot learning. The main idea is to acquire prior knowledge from a set of training tasks, which is then used to perform (few-shot) test tasks. Most existing work assumes that both training and test tasks are drawn from the same distribution, and a large amount of labeled data is available in the training tasks. This is a very strong assumption which restricts the usage of meta-learning strategies in the real world where ample training tasks following the same distribution as test tasks may not be available. In this paper, we propose a novel meta-learning paradigm wherein a few-shot learning model is learnt, which simultaneously overcomes domain shift between the train and test tasks via adversarial domain adaptation. We demonstrate the efficacy the proposed method through extensive experiments.

Keywords: Meta-Learning, Few-Shot Learning, Domain Adaptation

TL;DR: Meta Learning for Few Shot learning assumes that training tasks and test tasks are drawn from the same distribution. What do you do if they are not? Meta Learning with task-level Domain Adaptation!

Data: [MNIST-M](https://paperswithcode.com/dataset/mnist-m), [Office-Home](https://paperswithcode.com/dataset/office-home)