The objective of MAML is to provide a good initialization of a model's parameters in order to achieve optimal fast learning on a new task with fewer gradient steps. It also attempts to avoid overfitting scenarios, which happens while training a neural network with less data architecture. The following diagram is a representation of MAML:
As we can see in the preceding diagram, θ is the model's parameter and the bold black line is the meta-learning phase. Let's assume that we have three different new tasks and a gradient step is taken for each task (the gray lines with the arrowheads). We can see that the ...
