Transfer Learning

machine-learning

Definition

Transfer Learning

Transfer learning is a paradigm where knowledge extracted from a source task $T_S$ in a source domain $D_S$ is applied to a different but related target task $T_T$ in a target domain $D_T$. Formally, it aims to improve the learning of the target predictive function $f_T(\cdot )$ by leveraging parameters, features, or data distributions from $D_S$ and $T_S$.

Methodological Benefits

Transfer learning is particularly effective when $D_S\ne D_T$ (e.g., training on synthetic images and testing on real-world photos) or when labelled data is scarce in the target domain. By utilising a pre-trained model as a starting point, it significantly reduces the computational resources and amount of target data required to achieve high performance.