Optimal Bayesian Transfer Learning for Classifying Multivariate Gaussian Observations

The fundamental assumption to guarantee the generalizability of traditional machine learning methods is that there are sufficient data for training, assumed to be independently sampled with the identical underlying distribution as the future data to test. However, in many real-world applications involving complex systems, we may not have sufficient data or the collected data may manifest heterogeneity. In either case, the cross-cutting question is when and how we may “transfer” the “surrogate” knowledge and data from well-characterized systems (source) to a target system of interest, to improve the predictive power of machine learning methods for the target.