Jethro's Braindump

Meta Learning

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 \(\mathcal{T}\):

  1. Inner loop:

\begin{equation} \theta_i’ = \theta - \alpha \nabla_\theta L_{\mathcal{T}}(f_\theta) \end{equation}

  1. Outer Loop:

\begin{equation} \theta=\theta-\beta \nabla_{\theta} \sum_{\mathcal{T}_{i} \sim p(\mathcal{T})} \mathcal{L}_{\mathcal{J}_{i}}\left(f_{\theta_{i}^{\prime}}\right) \end{equation}