A Bayesian approach to (online) transfer learning: Theory and algorithms

• DOI: 10.1016/j.artint.2023.103991
• Publication Date: 2023-08-11T01:16:32Z
• Authors: Xuetong Wu; Jonathan H. Manton; Uwe Aickelin; Jingge Zhu
• Keywords: Inductive transfer; Transfer of learning; Instance-based learning; Negative transfer

Transfer learning is a machine paradigm where knowledge from one problem utilized to solve new but related problem. While conceivable that task could help task, if not executed properly, transfer algorithms can impair the performance instead of improving it – commonly known as negative transfer. In this paper, we use parametric statistical model study Bayesian perspective. Specifically, three variants problems, instantaneous, online, and time-variant learning. We define an appropriate objective function for each provide either exact expressions or upper bounds on using information-theoretic quantities, which allow simple explicit characterizations when sample size becomes large. Furthermore, examples show derived are accurate even small sizes. The obtained give valuable insights into effect prior learning, at least with respect our formulation particular, formally characterize conditions under occurs. Lastly, devise several (online) amenable practical implementations, some do require assumption. demonstrate effectiveness real data sets, focusing primarily source target have strong similarities.