A turbo for machine learning

MLOps transfers DevOps principles to machine learning. This promotes cooperation and accelerates projects.

Interest in the use of artificial intelligence (AI) and machine learning (ML) is unbroken. According to the “Machine Learning 2021” study by IDG Research Services, around two thirds of companies in Germany were already dealing with this topic in 2021. In the meantime, this proportion is likely to have increased further. As a result, concepts that aim to accelerate and optimize AI projects are gaining in importance. The Machine Learning Operations (MLOps) approach has recently attracted particular interest. “MLOps is an extension of the DevOps methodology to include resources from the fields of machine learning and data science in the DevOps ecology on an equal footing,” explains Christoph Nützel, Head of Technology and Data at Futurice, a digitization specialist who uses agile methods and MLOps himself uses.

A core capability of MLOps is the management of machine learning models throughout their lifecycle, including deployment, monitoring and governance, adds Joseph George, vice president of product management at BMC Software. “Machine learning models are treated and managed like software in terms of lifecycle management and development.”

An important function of MLOps is to create ML models that can be used multiple times. This has a favorable effect on development costs and reduces the time until ML applications are ready for use. A central element is also the automation of development and deployment steps. This applies to integration, testing, implementation and management of the necessary infrastructure.

The status quo of MLOps usage

(What: BARC)

In addition, developers and data scientists can better focus on core tasks such as developing machine learning algorithms when routine tasks are performed automatically – from validating to updating ML models. “MLOps also facilitates quality assurance through simplified debugging and makes models and their performance interpretable,” says Srikumar Ramanathan, Chief Solutions Officer at Mphasis, a provider of solutions and services for digitization projects. (See also interview.)

At the beginning of the journey

But the fact that MLOps offers advantages seems to be spreading slowly in German companies. Because according to the study “Driving Innovation with AI. Getting Ahead with DataOps and MLOps” by the Würzburg market research and consulting company BARC (Business Application Research Center), most companies in Germany are still at the beginning of their journey towards machine learning operations. According to this, only about half of the companies that have started ML projects use this approach and the complementary methodology DevOps (Development Operations).

According to BARC, this is surprising. At the same time, 97 percent of those surveyed stated that MLOps and DevOps lead to significant improvements in projects, for example in relation to the degree of automation of processes, faster time-to-market and better cooperation between project participants. “DataOps is aimed at realizing a manageable, maintainable and automated flow of quality-assured data to data products,” explains Alexander Rode, Data & Analytics Analyst at BARC. “And MLOps also addresses the special requirements regarding the development, deployment and maintenance of ML models, which are also data products.”

According to Jens Beier, Business Area Manager Business Applications & Data Analytics at Axians, one reason for the hesitant attitude towards MLOps is the fear of the IT environment becoming frayed: “Companies often use standard software tools such as SAP or Salesforce to have a homogeneous environment.” These platforms would often provide the context for AI and machine learning projects, despite the resulting possible friction losses.

The benefits of using MLOps

(What: BARC)

No silver bullet

As a first step, users of AI and machine learning should be clear about what goals they actually want to achieve with these technologies. According to Jens Beier, both are almost always embedded in an existing company or process context, for example in areas such as sales, production or service. “AI is not there as a miracle weapon, but rather optimizes and automates existing processes.”

It is therefore essential to check before introducing AI, ML and the corresponding “Ops frameworks” whether conventional approaches are not sufficient: “Often it turns out that a specific problem requires a different or simpler solution than initially thought”. , emphasizes Christoph Nützel from Futurice.

Another potential stumbling block is the database: “ML projects are not geared towards the respective use case, but towards the existing data,” Nützel continues. This fact often cannot be reconciled with conventional working methods. Therefore, it is necessary that developers, IT professionals, subject matter experts and data scientists find ways to collaborate and communicate effectively.

The MLOps process

Not only the human factor plays a role in the development and operationalization of machine learning models and corresponding applications. In the first step, it is necessary to define the problem that a user wants to address with the help of ML, such as improving quality assurance in a production environment. In addition, criteria must be defined that prove the success of the use of machine learning. This can be a percentage reduction in defective products in manufacturing.

The search for suitable data and its preparation plays a central role. “One reason ML projects fail is poor data sanitization,” states Joseph George of BMC Software.

This step provides for the division into data sets for training, testing and validation and is carried out by data scientists.

MLOps vs. AIOps: Services and Solutions

(What: Neptune AI Blog)

Rapid deployment through automation

This is followed by the training of machine learning models. Usually different algorithms are used. The model that presents itself as the most suitable variant after the completion of this work is evaluated and validated. One criterion is the ability of an ML model to make predictions. At this stage, automation capabilities are particularly important to keep the time between training and validation short. Short means that an ML model should be available after a few days.

The last steps are the implementation of the model in the operational environment and the monitoring. Continuous verification is important because machine learning is not usually used in static environments. Rather, it can make sense to modify a model, for example when new or supplementary data is available. This means that it is an iterative procedure with constant adjustments and optimizations. These tasks can also be automated as part of an MLOps approach.

Open source, cloud, commercial?

Companies looking to adopt MLOps have several options. “In general, MLOps and DevOps as process-oriented concepts can be implemented equally with open source and commercial tool stacks,” says Alexander Rode from BARC. According to the consulting firm’s study mentioned above, users of commercial tools are less likely to have problems with complexity when operationalizing machine learning models than users of open source tools. In general, Rode advises experimenting with different MLOps tools. Users should take current and future requirements into account.

One of the best-known open source tools is MFlows, a platform that allows users to manage the lifecycle of machine learning models, including testing and implementation. Other open source platforms include Google’s Kubeflow, Netflix’s Metaflow and MLReefs. In addition, MLOps experts can draw on a wide range of tools for special tasks. Examples include MLRun for developing and implementing ML models and AutoKeras, a library for Automated Machine Learning (AutoML). Another solution uses AutoML: H2O for the optimization of machine learning processes.

It is certainly a challenge for companies to put together an MLOps environment from many components. Especially when there is a lack of experts, such as data scientists and machine learning specialists. However, according to the German IT consultancy Viadee, Kubeflow is an open source platform that has the potential to become the dominant MLOps solution.

Another option are platforms and frameworks offered by providers such as BMC Software, Mphasis or Data Robot. They sometimes provide functions that go beyond MLOps. This is the case, for example, with BMC Helix Operations Management with AIOps. One component is operations management, another is AI-supported analysis and automation functions for IT operations.

Mphasis’ Pace-ML, on the other hand, combines model deployment pipelines and monitoring of ML models on one platform. According to the provider, this simplifies version control and maintenance of the models. Well-known IT companies such as IBM and HPE have also added machine learning operations to platforms such as HPE Ezmeral and IBM Cloud Pak for Data. It is helpful that users can implement some of the solutions either in a public cloud or in their own data center.

Another option is cloud-based MLOps platforms from providers such as AWS, Microsoft (Azure) and Google (Google Cloud Platform, GCP). They complement AI and machine learning services from these cloud service providers. For companies that already use such services, it is therefore worth considering also booking the MLOps component from the relevant provider.

However, Joseph George from BMC Software gives one thing to consider: “Large companies in particular are rarely pure cloud shops, but a mixture of multi-cloud and on-premises infrastructures as well as mainframe environments.” It is therefore important that cloud-based MLOps solutions can process data from such hybrid environments.

Christoph Nützel from Futurice, on the other hand, advises adapting the decision for MLOps tools to the IT strategy. The most important thing when choosing a solution is “how I can version data, ML models and features”. The answers give the direction. If necessary, users should fall back on the expertise of manufacturer-neutral consulting firms.


Source: com! professional by www.com-magazin.de.

*The article has been translated based on the content of com! professional by www.com-magazin.de. If there is any problem regarding the content, copyright, please leave a report below the article. We will try to process as quickly as possible to protect the rights of the author. Thank you very much!

*We just want readers to access information more quickly and easily with other multilingual content, instead of information only available in a certain language.

*We always respect the copyright of the content of the author and always include the original link of the source article.If the author disagrees, just leave the report below the article, the article will be edited or deleted at the request of the author. Thanks very much! Best regards!