Jethro's Braindump

Meta Learning

tags: Reinforcement Learning ⭐

Learning to learn: learn an update rule from related tasks

For example, tasks are related through a low-dimensional embedding.

Model-Agnostic Meta Learning (MAML)

Based on 2nd-order gradient descent:

2-stage gradient-based approach on batches of tasks $T$ :

Inner loop:

$θ_{i}^{'} = θ - α \nabla_{θ} L_{T} (f_{θ})$

Outer Loop:

$θ = θ - β \nabla_{θ} \sum_{T_{i} \sim p (T)} L_{J_{i}} (f_{θ_{i}^{'}})$

Resources

ICML 2019 Meta-learning Tutorial

Links to this note