Distributed system and method for performing machine learning on data records

Abstract

The invention provides a distributed system for performing machine learning on data record and a method thereof. The system comprises multiple computing devices of which each performs the same data flow computation about machine learning on its own data record; and a parameter server which is used for maintaining the parameters of a machine learning model, wherein when data flow computation for training the machine learning model is performed, the computing devices use the parameters acquired from the parameter server to perform the same operation about machine learning model training on the respective data record, and the parameter server updates the parameters according to the operation result of the computing devices; besides / or when data flow computation for estimation through the machine learning model is performed, the computing devices use the parameters acquired from the parameter server to perform the same operation about machine learning model estimation on the respective data record. Therefore, homogenization between the computing devices can be realized and the network transmission overhead can be reduced.

Description

technical field [0001] Exemplary embodiments of the present invention relate generally to the field of artificial intelligence, and more particularly, to a distributed system for performing machine learning on data records and a method for performing machine learning on data records using the distributed system. Background technique [0002] With the rapid growth of data scale, machine learning is widely used in various fields to mine the value of data. However, in order to perform machine learning, the memory of general physical machines is far from enough. Therefore, in practice, it is often necessary to use a distributed machine learning platform to complete the training of machine learning models or corresponding predictions. [0003] In existing distributed machine learning systems (for example, in Google's deep learning framework TensorFlow), there are usually one or more control nodes, which are responsible for scheduling tasks and computing resources of other computi...

AI Technical Summary

Problems solved by technology

[0005] Exemplary embodiments of the present invention aim to overcome the shortcomings of existing distributed machine learning systems that have large network overhead and computational load when performing machine learning

Images