Meta Learning recipe, black-box adaptation, optimization based approaches

Table of Contents

Probabilistic formulation of meta-learning(recap)

General recipe of meta-learning algorithms

Black-box adaptation approaches

Optimization-based meta-learning

Probabilistic formulation of meta-learning (recap)

General recipe of meta learning algorithms

많이 사용하는 dataset으로는 Omniglot, Mini-Imagenet, CIFAR, CUB, CelebA 등이 있다.

Mechanistic view

Probabilistic view

General recipe - Meta learning algorithm을 design 하는 방법

1. $p(\phi | \mathcal{D}^{\text{tr}}_i, \theta)$의 형태를 결정한다.
2. $\theta$를 maximum likelihood objective 관점에서 $\mathcal{D}_{\text{meta-train}}$을 이용해 어떻게 optimize할지 결정한다.

Black-box adaptation