A fast balanced two-branch network with model uncertainty for long-tailed data

In the natural data sets of daily life, medicine, agriculture, transportation and so on, there are a large number of sample imbalance problems. If a multi-classification data set is considered to be "unbalanced" or "skewed", there will be an expected deviation in the training and learning process of deep learning neural networks. To this end, we propose a fast and balanced dual-channel adaptive convolutional neural network (FBDAN) to solve the difficult problem of training on imbalanced datasets. FBDAN aims to use the uncertainty of the model to evaluate the training process and construct an efficient sample selector. We experiment on multiple datasets, the results show that the proposed network is better at recognising long-tailed data.