Transfer Learning

Related Concepts

• AdaBoost

• Domain Adaptation

• Few-Shot Learning

• Fine-Tuning

• Importance Sampling

• Transferable Features

Definition

Transfer learning is a methodology in machine learning that exploits the representation and/or features previously learned from some other tasks to better learn a target task. Generally, transfer learning makes learning target task faster, and when the target task lacks training data, transfer learning improves performance. Formally, a domain $\mathcal {D}=(\mathcal {X},P(X))$ consists of both the input space $\mathcal {X}$ and a probability distribution P(X) where $X \in \mathcal {X}$. Given a domain $\mathcal {D}$, a task $\mathcal {T}=(\mathcal {Y},f(\cdot ))$ includes both the label space $\mathcal {Y}$ and an objective predictive function f(⋅). In this entry, we consider transfer learning in the form of transferring from a source domain and task ($\mathcal {D}_S,\mathcal {T}_S$) to a target domain and task ($\mathcal {D}_T,\mathcal {T}_T$)....