KI.NRW and Fraunhofer IAIS publish study on machine learning operations (MLOps)

© monsitj – stock.adobe.com / KI NRW

Basics, opportunities and challenges of using MLOps in companies

What is MLOps? And how is it used by companies? In a study, experts from KI.NRW and the MLOps team at the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS interviewed a total of 29 companies to understand where they stand in their MLOps journey. The result is a compact overview of the basics, opportunities and challenges of using MLOps, which provides a detailed inventory as well as valuable recommendations on how companies can take action.

MLOps is a paradigm for the development and operation of machine learning (ML) applications. It involves both technical and organizational aspects, for example which infrastructure, which processes, team skills and which tools are required in the company for the productive use of ML.

This study examines the use of MLOps practices in companies and the extent to which individual aspects of these practices are already being applied. In just under 40 pages, the authors present the theoretical foundations, identify existing challenges and then derive recommendations for action. “We want to enable companies to successfully put AI into productive practice,” explains KI.NRW Managing Director Dr. Christian Temath, co-author of the study.

Companies from various sectors were surveyed. The prerequisite was that they had already dealt with the topic of machine learning operations. A particular focus was placed on SMEs in order to investigate the status of development and the need for support in the introduction of MLOps. “In terms of methodology, the study is based on the MLOps cycle that a typical AI project goes through,” explains author Dr. Dennis Wegener, MLOps Team Lead at Fraunhofer IAIS. This cycle consists of six phases and ranges from (1) requirement analysis, planning and design, through (2) exploration and (3) transfer to professional software development, to (4) integration and (5) testing of the application and finally to (6) operation and monitoring.

All steps are comprehensively examined in the study and the results are illustrated with the help of numerous visualizations. The AI experts estimate that many of the companies involved in the study already have a very high level of MLOps development. At the same time, there is still room for improvement in some areas, such as the tracking of experiments. Data management and the availability of data also continue to be a major challenge.

Overall, it is clear that the use of MLOps in companies can be very productive. The study provides a compact and straightforward explanation of how to use them successfully.

The study is available to download for free.