Model training method, device and server based on active learning

Abstract

Embodiments of the present application provide a model training method, device, and server based on active learning, which can automatically retrieve training samples for each application scenario indicated by the started training task, without the need for developers to manually export data for model training. After the initial deep learning network model pre-matched with each application scenario labels the training samples, the active learning service is dispatched to actively screen the labeled training samples, without the need for developers to participate in labeling, thereby effectively reducing the number of sample labels. After the active screening samples are calibrated according to the user's calibration instructions, the training service associated with each application scenario can be flexibly scheduled to perform model training based on the calibrated active screening samples and released to the corresponding software application for each application scenario. Furthermore, training services of different application scenarios can be flexibly connected, and automatic labeling, training and business updates of the training process can be realized.

Description

technical field [0001] The present application relates to the technical field of deep learning, in particular, to a model training method, device and server based on active learning. Background technique [0002] Among related technologies, most artificial intelligence technologies are related applications of supervised learning technology in machine learning. The development process of a supervised learning model requires the use of sample data for model training, and then the trained supervised learning model is packaged into an application module for deployment to the application scenario. When there is a certain difference between the sample data collected in the application scenario and the sample data used in training, the supervised learning model will not be able to meet the needs of the application, and it is necessary to re-collect sample data and train again, which will lead to an increase in the cost of the application , the application launch is delayed, and th...

AI Technical Summary

Problems solved by technology

When there is a certain difference between the sample data collected in the application scenario and the sample data used in training, the supervised learning model will not be able to meet the needs of the application, and it is necessary to re-collect sample data and train again, which will lead to an increase in the cost of the application , the application launch is delayed, and the efficiency of re-collecting samples and training again is low, and algorithm engineers with professional knowledge are also required to participate

Application scenarios Due to business changes and condition changes, the accuracy of the supervised learning model cannot meet the application requirements

[0003] The inventors of the present application found that for the deep learning model based on supervised learning, there are subtle differences between the training data and the field data of the application scene, which may cause the model to fail to meet the requirements of the application. Data and retraining the model to improve the ability of the original model to adapt to the field environment. This method is complex to operate and has a low degree of automation. It requires the participation of professional developers, and the work of sample labeling is large. It is often necessary to update the field application system to use the retrained model, leading to low R&D efficiency and even affecting business delivery and launch

Images