Neural network training method and device suitable for long-tail distribution data set

Abstract

The invention provides a neural network training method and device suitable for a long-tail distribution data set. The neural network comprises a feature extraction network, a classifier and a category gradient reweighting network. The training method comprises the following steps: obtaining a training sample set; extracting, bythe feature extraction network, features from the training sample setto obtain features, classifying the features through a classifier, and establishing a loss function according to a classification result; calculating the gradient of each neuron in the feature extraction network in the training sample according to the loss function; and in the back propagation process of neural network training, calculating, by the class gradient reweighting network, the reweighting gradient weight of the training sample, and adjusting the gradients of the training samples belonging to different classes according to the reweighting gradient weight. Therefore, the method solvesa problem that the recognition accuracy of the neural network is reduced under the training data of the long-tail distribution, alleviates the overfitting phenomenon of the feature extraction network, and improves the recognition accuracy and robustness of the deep neural network under the long-tail distribution.

Description

technical field [0001] The present invention relates to the technical fields of artificial intelligence and deep learning, in particular to a neural network training method and device suitable for long-tail distribution data sets. Background technique [0002] With the rapid development of deep learning and neural networks, deep learning techniques are widely used in computer vision applications, such as object recognition, object detection, semantic segmentation, etc. Training a neural network often requires training data with sufficient data volume and balanced data distribution. However, collecting such training data often requires a lot of manpower and material resources. In the early target recognition algorithms based on neural networks, data sets with balanced data and small data volume were often used, such as MNIST and CIFAR. The former is training data for handwritten digit recognition, and the latter is training data for general object recognition, and its categ...

AI Technical Summary

Problems solved by technology

[0008] Most of the above three methods can alleviate the difficulty of neural network identification under the long-tail distribution, but they also have their own limitations, and at the same time, the analysis of the characteristics of the neural network itself is insufficient.

Images