In the traditional machine learning methods,it is assumed that the training and test data have an identical distribution.However,this assumption is not valid in many cases.In order to solve this problem,a non-parametric transfer learning algorithm named Multi-Step Bridged Refinement is proposed.In this algorithm,a series of intermediate models is constructed to bridge different domains,and the label propagation between neighboring mo-dels is performed,through which the discriminative information is transferred from the source domain into the target one.Experimental results show that the models with similar distribution contribute to smooth transfer and make the refinement results insensitive to the initial label,and that the proposed algorithm attains a classification accuracy higher than that from other algorithms.