Bayesian neural multi-source transfer learning

Abstract

Although the use of deep learning and neural networks techniques are gaining popularity, there remain a number of challenges when multiple sources of information and data need to be combined. Although transfer learning and data fusion methodologies try to address this challenge, they lack robust uncertainty quantification which is crucial for decision making. Bayesian inference provides a rigorous approach for uncertainty quantification in decision making. Uncertainty quantification using Bayesian inference takes into consideration uncertainty associated with model parameters, as well as, the uncertainty in combining multiple sources of data. In this paper, we present a Bayesian framework for transfer learning using neural networks that considers single and multiple sources of data. We use existence of prior distributions to define the dependency between different data sources in a multi-source Bayesian transfer learning framework. We use Markov Chain Monte-Carlo method to obtain samples from the posterior distribution that consider the knowledge from the source datasets as priors. The results show that the framework provides a robust probabilistic approach for decision making with accuracy that is similar to gradient-based learning methods. Moreover, the results are comparable to related machine learning methods used for transfer learning in the literature.