Transfer learning definition via domains and tasks

Summary

Transfer learning is defined in terms of two abstractions: domains and tasks. A domain D consists of a feature space X and a marginal probability distribution P(X) over the features. A task T consists of a label space Y and a predictive function f: X → Y that maps an input x to a label f(x). The task is learned from training data made of input–label pairs {x_i, y_i}. In transfer learning, there is a source domain D_S with a source task T_S and a (possibly different) target domain D_T with a target task T_T. The goal is to improve the target predictive function f_T(·) in the target domain by reusing knowledge from the source domain and source task. Transfer occurs when either the domains differ (D_S ≠ D_T) or the tasks differ (T_S ≠ T_T), enabling better learning efficiency on the target problem.

Key Points

• 1

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

• 2

  A task is defined by a label space Y and a predictive function f: X → Y learned from labeled pairs {x_i, y_i}.

• 3

  Transfer learning uses a source domain/task to improve a target predictive function when domains and/or tasks differ.