Meta-learning

Meta-learning is another exciting trend of research in deep learning. It goes beyond training on a huge dataset for a specific task, like traditional learning does. It tries to mimic a humans' learning process by leveraging a past experience that has been learned from a distribution of tasks. It can achieve good performance, even with just a handful of training samples. However, conventional deep learning methods are not able to do so. 

One big challenge in deep learning – training data

You might have seen the following plot, comparing the performance between deep learning and traditional machine learning algorithms given various amount of training data:

With only a small amount of training...