Information interaction method and system

Abstract

The application discloses an information interaction method and system. The information interaction method comprises: whether an exit command is received is determined; if not, current weight information is read from a first device and gradient information is calculated based on the read current weight information; and gradient information that is larger than or equal to a preset gradient threshold value is returned to the first device and thus the first device calculates weight information again according to the returned gradient information and uses the weight information that is larger than or equal to a preset weight threshold value as current weight information, and the operations are executed repeatedly until the exit command is received. According to the information interaction method and system, the gradient information that is larger than or equal to the preset gradient threshold value is returned to the first device by at least one second device and the weight information that is larger than or equal to the preset weight threshold value is provided for at least one second device by the first device, so that the communication load between the first device and the second device is reduced substantially and thus the resource consumption is reduced and the communication efficiency is improved.

Description

technical field [0001] The present application relates to the field of computer technology, in particular to an information interaction method and system. Background technique [0002] As the core technology of big data processing, machine learning can be widely used in various fields, especially for generating training models. Since the loss function variables (that is, weight information) involved in training the model may exceed hundreds of billions, the training data is very large, and there is no one-step solution to minimize the loss function. It usually takes multiple rounds of iterations to obtain an approximate optimal solution. [0003] Specifically, in machine learning, slave (also known as worker / worker, in a cluster environment, there may be multiple workers) calculates gradient information and sends all gradient information to master (also known as server / server ), then the master calculates the new weights according to the weight update formula and the gradie...

AI Technical Summary

Problems solved by technology

In this way, the master and the slave can obtain an approximate optimal solution through multiple rounds of iterations, and can obtain a better training model. However, during the iteration process, the amount of communication between the two is huge, which will inevitably increase resource consumption and reduce Communication efficiency, in addition, slave memory consumption is huge

Images