Hybrid neural network model for large-scale heterogeneous classification tasks in few-shot learning

Abstract

How to generalize and unify different few-shot learning tasks using neural network model is a difficult problem in the field of machine learning research. Aiming at the problem that the parameters of existing few-shot learning models cannot adapt with heterogeneous classification tasks, inspired by the human being recognition process, a hybrid neural network (HNN) model for large-scale heterogeneous classification tasks in few-shot learning is proposed. First, a meta-learning model is constructed, which uses a siamese graph convolutional network (SGCN) structure as bone network. The SGCN is trained by semi-supervised way with a small amount of incomplete labeled data. Then, random task slicing by group is performed according to the task size and meta-learning dimensions to ensure that the segmented task size matches the meta-learning model. Combined with the meta-learning model, a task discrimination network and object recognition network are constructed, to perform heterogeneous classification tasks while keeping the scale of HNN network parameters unchanged. Experimental results show that the HNN performs well under different datasets, and is suitable for large-scale heterogeneous tasks in few-shot learning without retraining.