Transfer learning definition via domains and tasks

Summary

Transfer learning is defined in terms of two related concepts: domains and tasks. A domain D consists of a feature space X and a marginal probability distribution P(X) over instances x in X. A task T includes a label space Y and an objective predictive function f that maps inputs from X to labels in Y, learned from training pairs (x_i, y_i). The goal is to use the learned predictive function to predict labels for new instances. When transferring, a source domain D_S and source task T_S are used to improve learning in a different target domain D_T and target task T_T. Transfer learning is specifically aimed at improving the target predictive function f_T(·) in D_T by leveraging knowledge from D_S and T_S, even when the domains differ (D_S ≠ D_T) and/or the tasks differ (T_S ≠ T_T).

Key Points

• 1

  A domain is characterized by a feature space X and a marginal distribution P(X).

• 2

  A task is characterized by a label space Y and a predictive function f: X → Y learned from labeled data.

• 3

  Transfer learning uses knowledge from a source domain/task to improve a target predictive function in a different target domain/task.