INDUCTIVE GENERALIZED ZERO-SHOT LEARNING BASED ON GAUSSIAN-MIXTURE-LIKE MAPPING

In the study of generalized zeroshot learning, the partial dependence of the classifier on the visible class and the information loss in the process of highdimensional to lowdimensional feature mapping are two main problems in traditional algorithms. In order to solve the problems, based on the idea of Gaussian mixture distribution model and combined with the design concept of common learning, this paper proposes a multi-channel structure. The structure can not only realize supervised common learning between channels by establishing channel learning rate differentiation, but also can fit the real distribution characteristics of the generated features through the calculation of multi-Gaussian distribution and enhance the feature mapping capability of the network in hidden space. In order to verify the multi-channel structure, this paper conducted a large number of experiments on three benchmark databases CUB, AWA2 and SUN. Harmonic index H has increased 1.4,1.56 and 0.47 for I-GZSL. It proves the effectiveness of the multi-channel structure in generalized zero shot learning.