A conversion system, method and electronic device for machine learning algorithm

Abstract

The disclosure provides a machine learning algorithm conversion system, method, and electronic equipment, which relate to the field of computer technology. The system includes: a programming interface layer, which is used to construct a data flow diagram of an original machine learning algorithm based on a preset data flow generation tool; The operator placement evaluation module is used to calculate the corresponding placement cost of each operator in the data flow graph when executed by different participants; the data flow graph split scheduling module is used to divide the data flow graph into multiple sub-parts according to the placement cost graph, and dispatch the subgraph to the target participant for execution; the compilation execution layer is used to compile the subgraph into a new data flow graph based on the greedy algorithm strategy, and generate the instructions of each operator in the new data flow graph to obtain the distributed Privacy Preserving Machine Learning Algorithms. The disclosure can effectively improve the problems of poor compatibility, high coupling degree and poor performance existing in the algorithm conversion process.

Description

technical field [0001] The present disclosure relates to the field of computer technology, in particular to a machine learning algorithm conversion system, method and electronic equipment. Background technique [0002] The various training data used to drive the development of AI models (such as Alpha Go, GPT-3) are often scattered in various institutions, so distributed privacy-preserving machine learning (also known as federated learning) can be used. To solve the data privacy protection problems encountered in geographically distributed data collection and model training. [0003] One data segmentation method in distributed privacy-preserving machine learning is vertical segmentation: data is distributed among different participants according to characteristics. For vertically split distributed privacy-preserving machine learning, the existing conversion method is to first implement a common serial version of the algorithm, and then transplant the algorithm to a distribu...

AI Technical Summary

Problems solved by technology

[0004] However, the above methods have the main problems of poor compatibility, high coupling and poor performance

Among them, poor compatibility refers to distributed privacy-preserving machine learning. As an extension of machine learning algorithms, it has poor compatibility with mainstream deep learning frameworks and is not convenient for developers to use.

A high degree of coupling means that machine learning algorithms and privacy-preserving computing protocols are tightly coupled in current distributed privacy protection scenarios. Almost every development requires careful analysis of the entire process of the original machine learning algorithm; Issues such as difficulty in iteration and poor scalability

Poor performance means that distributed privacy-preserving machine learning is generally 100 or even 1000 times slower than the original stand-alone machine learning algorithm, which brings difficulties to the actual implementation of distributed privacy-preserving machine learning

Images