A deep learning method with stable performance

Abstract

The invention provides a deep learning method with stable performance, which belongs to the technical field of deep learning. The method obtains a labeled image dataset as a training dataset, and constructs a deep learning network composed of a deep feature extractor and a classifier; randomly selects a batch of samples from the training dataset and inputs it into the network, wherein the deep feature extraction The device outputs the original features of the batch of samples and undergoes random Fourier feature transformation to obtain the corresponding random Fourier representation matrix; uses the random Fourier representation matrix to detect the independence of the original features, and trains to obtain the corresponding weights of each sample; After re-weighting the predicted loss value, the final training loss value is obtained and the network parameters are updated until the end of the deep learning network training. The present invention can be applied to a picture recognition system and a target detection system with cross-distribution problems, and solves the problem of accuracy drop caused by the deviation of the distribution of training data and test data.

Description

technical field [0001] The invention belongs to the technical fields of image recognition, object detection and the like, and particularly proposes a deep learning method with stable performance. Background technique [0002] At present, deep learning has made unprecedented progress in many research fields, especially in the field of computer vision (such as image recognition, object detection and other technical fields). For example, residual convolutional networks can greatly improve the recognition accuracy of images by computer vision recognition systems. Based on The convolutional network of the region can greatly improve the accuracy of the object detection system and so on. Many deep learning-based computer vision techniques have far surpassed traditional methods. [0003] However, most of the current machine learning and deep learning algorithms assume that the training data and test data are independent and identically distributed, and have achieved very good resul...

AI Technical Summary

Problems solved by technology

The present invention can be applied to image recognition systems and target detection systems that have cross-distribution problems, and solves the problem of accuracy drop caused by distribution offsets between training data and test data

Images