Method and system for predicting generalization error of image recognition model based on non-check set

Abstract

The invention discloses a method and a system for predicting generalization errors of an image recognition model based on a non-check set, and belongs to the field of deep learning optimization and generalization. The method comprises the steps: randomly sampling K groups of training pictures after each training round, and calculating the parameter updating amount of the image recognition model corresponding to the K groups of training pictures through a model optimizer; utilizing the parameter updating amount to obtain corresponding K updated models, and recording the output of the K updatedmodels to each training picture; calculating an output variance value of each training picture, and normalizing the variance values by using the output module length to obtain an output relative variance; and outputting the variation trend of the generalization errors of the relative variance prediction image recognition model in the training process. Therefore, the method does not need to use a verification set, so that all training samples can be put into training, and better generalization performance is obtained; in addition, only one round of neural network needs to be trained in the process, and energy and hardware losses caused by multiple times of training are reduced.

Description

technical field [0001] The invention belongs to the field of deep learning optimization and generalization, and more specifically relates to a method and system for predicting the generalization error of an image recognition model based on a non-verification set. Background technique [0002] As a current research hotspot in artificial intelligence, machine learning is often used to mine potential relationships between data. In recent years, data-driven machine learning algorithms have achieved outstanding results in various fields such as biology, medicine, finance, and military affairs. With the improvement of data and computing power, deep learning, as a machine learning algorithm that can process images well, has become a current research hotspot and is widely used in various industries. [0003] Although deep learning has a good performance in the task of image recognition, there are still many problems to be solved and studied urgently. The neural network model used ...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above defects or improvement needs of the prior art, the present invention provides a method and system for predicting the generalization error of an image recognition model based on a non-verification set, thereby solving the problem of using a verification set to predict errors in the training process of an existing image recognition model. The technical problems of multiple training costs and inaccurate predictions in generalization performance

Images