Automated deployment of machine learning applications to the cloud

The use of machine learning (ML) as a key technology in artificial intelligence (AI) is becoming more and more important in the increasing digitalization of business processes. However, the majority of the development effort of ML applications is not related to the programming of the ML model, but to the creation of the server structure, which is responsible for a highly available and error-free productive operation of the ML application. The creation of such a server structure by the developers is time-consuming and complicated, because extensive configurations have to be made. Besides the creation of the server structure, it is also useful not to put new ML application versions directly into production, but to observe the behavior of the ML application with respect to unknown data for quality assurance. For example, the error rate as well as the CPU and RAM consumption should be checked. The goal of this thesis is to collect requirements for a suitable server structure and an automation mechanism that generates this server structure, deploys the ML application and allows to observe the behavior of a new ML application version based on real-time user data. For this purpose, a systematic literature review is conducted to investigate how the behavior of ML applications can be analyzed under the influence of real-time user data before their productive operation. Subsequently, in the context of the requirements analysis, a target-performance analysis is carried out in the department of a management consulting company in the automotive sector. Together with the results of the literature research, a list of user stories for the automation tool is determined and prioritized. The automation tool is implemented in the form of a Python console application that enables the desired functionality by using IaC (Infrastructure as code) and the AWS (Amazon Web Services) SDK in the cloud. The automation tool is finally evaluated in the department. The ten participants independently carry out predefined usage scenarios and then evaluate the tool using a questionnaire developed on the basis of the TAM model. The results of the evaluation are predominantly positive and the constructive feedback of the participants includes numerous interesting comments on possible adaptions and extensions of the automation tool.

Level thesis:degree_level

master

Grantor dc:publisher

Universität Heidelberg

Year

2020

Author dc:creator

• Radeck, Leon

Contributors dc:contributor

• Paech, Barbara

Repository record source_url

http://www.ub.uni-heidelberg.de/archiv/28982

OAI identifier oai:identifier

oai:archiv.ub.uni-heidelberg.de:28982