Abstract

Untitled

목적: 적은 수의 학습 sample 만 사용해도 새로운 learning task를 해결하는 것

Introduction

Meta-learning

Optimization-based approach

Few shot learning

Meta-learning vs Multi-task learning

Model-Agnostic Meat-Learning

Task function T = {L(x₁; a₁; ... ; xH; aH); q(x₁); q(xₜ₊₁|xₜ; aₜ); H}

<aside> 💬 L: loss function

q(x₁): initial distribution

q(xₜ₊₁|xₜ; aₜ): transition distribution

H: episode leangth (i.i.d & H = 1 → supervised learning)

</aside>

Untitled

Untitled

<aside> 💡 key idea 어떠한 internal representaions 보다 더 빨리 adapt할 수 있는 representation이 존재할 것이다.

</aside>