Adaptive Knowledge Transfer based on Transfer Neural Kernel Network

Transfer agents are widely used in the challenging problems where knowledge is cross-used among different tasks. One popular research approach is to design a transfer kernel that controls the strength of knowledge transfer based on the similarity of tasks. In this paper, we propose a Transfer Neural Kernel Network (TNKN), which enables flexible modeling of the task similarity. The proposed TNKN is constructed by compositions of primitive kernels and represented by a neural network. Two coupled compositional kernel structures are used to characterize data covariance, one for the intra-task data covariance and another for the inter-task one. A sufficient condition that validates the transfer agent using TNKN for any data is given. This condition also discloses the relationship of the two compositional kernel structures, and can be used as a constraint in the agent learning. Since the overall architecture of TNKN is differentiable, the learning of the transfer agent using TNKN is end-to-end trainable with gradient-based optimization. Extensive experiments on various real-world datasets demonstrate the transfer effectiveness of TNKN.