[Paper] A survey of transfer learnings

“A survey of transfer learning”이란 논문에 대한 리뷰입니다.

원문은 링크에서 확인할 수 있습니다.

Terminology

Feature space Χ / Label space Y Predictive Function f(∙)=P(Y│X) Domain D={Χ,P(Χ)} Task Τ={Y,f(∙)}

	Source	Target
Domain	$X_{S},P(X_{S})$	$X_{T},P(X_{T})$
Domain data	$D_S={(x_{S1},y_{S1} ),…,} $	D_T={(x_{T1},y_{T1}),…,}
Task	$Y_{S},f_{S}(∙)$	$Y_{T},f_{T}(∙)$

Definition

improving target predictive function using source domain data & target domain data.

• X, P(X)가 달라서 생기는 문제로 귀결된다.
• Domain Adaptation이라는 것과 혼용되지만 이건 source domain을 target domain과 비슷하게 만드는 방식이다.

Taxonomy

|Feature Space Homogeneous| Heterogeneous| |:—:|:—:| |Predictive Function|Mismatch in the conditional prob| |Label Space|Mismatch in the class space| |P(Y) | Caused by labelled and unlabelled|

Bias

||| |:—:|:—:| |Frequency feature bias |P(X_S )≠P(X_T )| |Context Feature bias |$P(Y_{S}|X_{S})≠P(Y_{T}|X_{T}})$|

Generic Solution

||| |:—:|:—:| |Homogeneous | Correct marginal and/or conditional| |Heterogeneous |(Same domain distribution) Align Input space| || (Different domain distribution) domain adaptation|

General Strategy

||| |:—:|:—:| |Information transfer|(Through Instances) reweight source domain to correct marginal
-> Conditional is the same| |Information transfer|(Through features) (Asymmetric transformation) Transform features through reweighting| |Information transfer| (Symmetric transformation)Find common latent feature space| |Information transfer|(Through Parameter, Ensemble learning) multiple source learners | |Information transfer|Transfer based on some defined relationship(least)|

Generic solution + General strategy  Methodology