Human-in-loop machine learning application development method and system

Abstract

The invention relates to a human-in-loop machine learning application development method and system, a machine learning application development process is regarded as a control process, a data flow isa signal flow, and according to a negative feedback adjustment principle in a control theory, a negative feedback loop comprising an online stage and an offline stage and three artificial assistantsis designed. Data of the on-line loop can enrich basic data samples, fault feedback of the on-line loop is matched with manual processing to position the off-line data or the model stage, then the problem is solved, and finally the model data flow is uploaded to the on-line state to form a continuous iterative development process. The invention provides a solution for developing a complete set ofprocesses of machine learning application. According to the method, based on the negative feedback adjustment principle in the control theory, manual experience and knowledge are applied to loops of all development stages at a relatively low cost, iterative and incremental development is supported, and the development quality of all components and stages is improved, so that the performance of thewhole application is improved.

Description

technical field [0001] The invention relates to a human-in-the-loop machine learning application development method and system, belonging to the field of computer technology applications. Background technique [0002] At present, with the continuous development of data collection and storage capabilities, after possessing massive data resources and distributed parallel high-performance computing capabilities, it is necessary to write algorithms to calculate and mine value from data. The algorithms here generally refer to machine learning. Generally speaking, machine learning applications are mainly divided into the following three parts: 1. Data level, including feature engineering such as feature cleaning, feature conversion, and feature selection; 2. Model training level, including model algorithms, model tuning, model evaluation, etc. ; Three, the application level, model deployment, online and service. Machine learning is a high-cost, high-threshold work that requires e...

AI Technical Summary

Problems solved by technology

[0003] The following problems exist in the development of a machine learning application: 1) There are some gaps between the various stages of the machine learning process, which are not coherent; 2) The cost and threshold of machine learning itself are high, and the specific development process requires the expertise of algorithm practitioners Experience and wisdom; 3) The machine learning process itself is a stage of continuous training, continuous learning, and continuous optimization, which requires continuous improvement

In the industry, Alibaba proposed the maxcompute computing platform based on its self-developed storage and computing technologies to connect the entire process. Continuous and iterative loops are still inconvenient for application iteration and upgrades. At the same time, the existing platforms do not support the diversity of multiple machine learning frameworks and programming languages ​​enough

Images