Few-shot learning with deep balanced network and acceleration strategy

Deep networks are widely used in few-shot learning methods, but deep networks suffer from large-scale network parameters and computational effort. Aiming at the above problems, we present a novel few-shot learning method with deep balanced network and acceleration strategy. Firstly, a series of simple linear operations are applied to few original features to obtain the more features. More features are obtained with fewer parameters, thus reducing the network parameters and computational effort. Then the local cross-channel interaction mechanism without dimensionality reduction is used to further improve the performance with nearly no increase in parameters and computational effort, so as to obtain a deep balanced network to balance performance, parameters, and computational effort. Finally, an acceleration strategy is designed to solve the problem that the gradient update in the deep network takes a tremendous amount of time in new tasks, speeding up the adaptation process. The experimental results of traditional and fine-grained image classification show that the few-shot learning method with deep balanced network can achieve or even exceed the classification accuracy of some existing methods with fewer network parameters and computational effort. The cross-domain experiments further demonstrate the advantages of the method above the domain shift. Simultaneously, the time required for classification in new tasks can be significantly decreased by using the acceleration strategy.