B1307
Title: Addressing the validity of information borrowing in transfer learning
Authors: Anjishnu Banerjee - Medical College of Wisconsin (United States) [presenting]
Abstract: The aim is to address the methodological constraints around the basic premise of information borrowing in Bayesian versions of transfer learning. In general, distributed inference, where inference is made from piece-wise data, borrowing of information from related but mixed domain models, and cases when borrowing of information occurs in related but externally differentiated models (through model propagation or convolution) are considered. Specific inferential methods are discussed to incorporate pre-trained knowledge and external data. Enabling external data information borrowing allows one to gain efficiency without having to "reinvent the wheel". In contrast, hierarchical and adaptive structures allow deviations from information gleaned from external data. While focusing on Bayesian learning, the investigations considered are generalizable to other contexts. A novel methodology and theoretical considerations are presented, which enable inferential probabilistic guarantees and efficient model computation using both simulated and real examples.
