Homogeneous transfer learning for supporting pervasive edge applications

Edge computing is a paradigm which refers to the use of a range of devices and computational capabilities close to end users. The main idea is to process data closer to the location they are being generated, enabling processing with low latency leading to a model that is appropriate to support real time inference. Transfer learning and its usage in edge computing have recently received significant attention by the research community. Many research efforts focus on the improvement of transfer learning and/or exploit it in the edge setting. A major limitation of these efforts is the failure to consider the diverse statistical distributions representing each dimension of a dataset. In this paper, we provide a novel mechanism for identifying multidimensional distribution-based similarity on the available datasets and performing transfer learning according to this similarity. Additionally, our proposed model, Distribution Based Transfer Learning (DBTL), effectively addresses the uncertainty inherent in these setups, which influences the decision-making process for determining the optimal transfer learning and training procedures. We elaborate on the use of fuzzy logic to deliver an uncertainty-driven decision making model and present the outcomes of our experimental evaluation. The ultimate goal is to reveal the advantages of the proposed solution and provide its applicability in real setups.