Few samples learning based on granular neural networks

In system modeling, traditional machine learning methods aim to make a model’s output fit the real output as well as possible. However, sometimes they fail to reach the goal, especially when the quantity of samples is small (which often leads to the occurrence of overfitting). Therefore, researchers start to explore new approaches for system modeling with few samples. An effective alternative to solve this problem is to reduce the fitting expectation and make predictions from a certain range of samples. Granular computing techniques simulate human’s thinking rules at a higher level and thus can be applied to few samples learning (FSL). In this paper, we realize granular neural network (GNN) modeling with few samples. A conceptually simple yet powerful method to learn some coarse-grained information from few samples is proposed. The output of the GNN model is fuzzy (represented by information granules). It can predict a new sample’s output with a rough range which makes the model own stronger robustness especially when the quantity of training samples is small. The precondition is that it is not necessary to accurately determine the explicit output of the samples. In the experiment, we compare the coverage of the models on the same test set in which the models are built using different percentage of training samples. The results show that the model built on few samples can also achieve good performance.