Computational Graph in Model Agnostic Meta Learning

I found meta-learning amusing for the start especially this paper Model Agnostic Meta Learning.

Meta Learning Idea

It’s fancy name for training neural-networks in such a way that, they can generalize to a new task with few training examples. For example, training a classification algorithm on your custom-usecase where you have only 5 training examples available for each class.

Internet is full of resources for meta-learning. But I am interseted in discussion of a specific model architecture called MAML. More specifically, I will visualize the computationl graph in this architecture because it involves second-order optimization.

Dissecting MAML Algorithm

WIP. I have drawn out things. But need to validate stuffs which haven’t got time to.