Robust Bayesian Transfer Learning Between Kalman Filters

Bayesian transfer learning typically requires complete specification of the stochastic dependence between source and target domains. Fully probabilistic design-based Bayesian transfer learning-which transfers source knowledge in the form of a probability distribution-obviates these restrictive assumptions. However, this approach has suffered from negative transfer when the source knowledge is imprecise. We propose a scale variable relaxation to transfer all source moments successfully, achieving robust transfer (i.e. rejection of imprecise source knowledge). A recursive algorithm is recovered via local variational Bayes approximation. The solution offers positive transfer of precise source knowledge, while rejecting it when imprecise. Experiments show that the technique is competitive with or equivalent to alternative methods.