Research On Graph Neural Networks For Small-scale Datasets

With the development of hardware devices(e.g.GPU)and the publication of large-scale datasets,deep learning achieves good performance in the computer vision.Deep neural networks stack neurons layer by layer and contain lots of trainable parameters.To train a deep neural network,a large number of annotated data are used to perform the forward propagation and backward propagation,which is to update the parameters of the network.However,the annotation of a large-scale dataset requires manual participation.The data collection and the data annotation are both difficult in some tasks,which results in high costs.In addition,the data annotation in some tasks requires professional knowledge,which further increases the difficulty of the data annotation.Thus,some tasks use limited annotated data to train a deep neural network.However,deep neural networks cannot generalize well when the models are trained with limited training data.Therefore,the research on deep learning for small-scale datasets is important.To improve the performance of deep neural networks on small-scale datasets,the proposed method first analyzes the internal relation in limited data,and then constructs the relational graph.Finally,it uses the graph neural networks to share the information,which deals with the lack of information on small-scale datasets.This paper studies the visual reasoning task,fine-grained classification task,and multi-label classification task.For the visual reasoning task,the proposed method constructs the relationships among objects with the guidance of queries.Based on the relational graph,the proposed method shares information between the objects,and then predicts the referent correctly.Since the training images of fine-grained datasets are limited,the proposed method constructs a category-attribute relational graph.By sharing the information between categories and attributes,the model obtains the important areas of the images for better classification.For the multi-label classification task,each image has multiple labels and the annotation may be incomplete.The proposed method constructs a relational graph to represent the relationships between categories.By sharing the information of related categories,the accuracy of the model is further improved.In the experiment,the proposed method is evaluated on the datasets of the visual reasoning task,the fine-grained classification task,and the multi-label classification task,respectively.The experimental results of the proposed method are better than other methods.It shows that sharing the information by graph neural networks improves the performance of deep neural networks on small-scale datasets.