How to solve this ?

There are few different approaches and are briefly described here.

Transfer Learning

It is fairly simple algorithm and easy to implement. Unlike the typical case of transfer learning, there is a subtle difference in how we handle train and validation as this involves training and validation through tasks , not just with some batch of images and labels.

Transfer Learning

Metric Learning

In this approach, we will learn a metric that differentiate between the classes in the support set. It is quite popular and it is the first kind of approach towards solving this problem.

Metric Learning

Data Augmentation

It is basically learning to augment. Most of it is based on GANs. DAGAN is the popular one.

Data Augmentation

Meta Learning

It is like learning to learn. The basic algorithms in these area are MAML. There are other variations like MAML++, Meta-SGD etc.,.

Meta Learning