To accelerate AI development, how can companies use MLOps to improve production efficiency?

When companies deploy artificial intelligence and build machine learning projects for the first time, they often focus on theory. So is there a model that can provide the necessary results? If so, how do we build and train such a model?

According to IDC data, it takes more than 9 months on average to deploy artificial intelligence or machine learning solutions. Mainly because the tools data scientists use to build these proofs of concept often don’t translate well to production systems. IDC analyst Sriram Subramanian said: "We call the time required for the R&D process 'model speed', that is, how long it takes from start to finish."

Enterprises can use MLOps to solve the above problems. MLOps (Machine Learning Operations) is a set of best practices, frameworks and tools that can help enterprises manage data, models, deployment, monitoring, and other aspects that use theoretical concepts to validate AI systems and make them effective.

Subramanian further explained, “MLOps reduces model speed to weeks—sometimes even days, just like using DevOps to speed up the average time to build an application, which is why you need MLOps.” Enterprises adopt MLOps can build more models, innovate faster, and cope with more usage scenarios. “The value proposition of MLOps is clear.”

According to IDC, 60% of enterprises will use MLOps to implement their machine learning workflows by 2024. Subramanian said that when they surveyed respondents about the challenges of adopting AI and machine learning, one of the top barriers was the lack of MLOps, second only to cost.

In this article, we examine what MLOps is, how it has evolved, and what organizations need to accomplish and keep in mind to make the most of this emerging approach to AI operations.

The Evolution of MLOps

A few years ago, when Eugenio Zuccarelli first started building machine learning projects, MLOps was just a set of best practices. Since then, Zuccarelli has worked on AI projects at several companies, including those in healthcare and financial services, and he has seen MLOps begin to evolve over time to include a variety of tools and platforms.

Today, MLOps provides a fairly powerful framework for operating artificial intelligence, said Zuccarelli, now an innovation data scientist at CVS Health, citing a previous project he worked on to create a system that could Applications for predicting adverse outcomes such as readmission or disease progression.

“We are exploring data sets and models and communicating with doctors to find out the characteristics of the best models. But for these models to be truly useful, users need to actually use these models.”

This means building a reliable, fast and stable mobile application, with a machine learning system on the back end connected through an API. “Without MLOps, we wouldn’t be able to ensure this,” he said. His team used the H2O MLOps platform and other tools to create health dashboards for the model. “You definitely don’t want major changes to the model, and you don’t want to introduce bias. The health dashboard allows us to understand whether the system has changed.”

Updates to production systems can also be made by using the MLOps platform. He said: "It is very difficult to swap out files without stopping application work. MLOps can swap out the system while production is in progress with minimal system impact."

He said, MLOps As the platform matures, it will speed up the entire model development process, because companies will not have to reinvent the framework for every project. Data pipeline management capabilities are also critical to AI implementation.

"If we have multiple data sources that need to communicate with each other, this is where MLOps comes into play. You want all the data flowing into the machine learning model to be consistent and of high quality. Just like that sentence As the saying goes, garbage in, garbage out. If the model's information is poor, then the prediction itself will be poor."

The foundation of MLOps: an ever-changing target

But don't think , simply because there are so many platforms and tools available, the core principles of MLOps are ignored. Businesses new to MLOps should remember that at its core, MLOps is about creating a strong connection between data science and data engineering.

Zuccarelli said: "To ensure the success of MLOps projects, you need data engineers and data scientists to be working on the same team."

In addition, prevent bias, ensure transparency, and provide accountability Interpretability, as well as the tools necessary to support an ethics platform, are still under development. "There is definitely a lot of work that needs to be done in this area because this is a very new area."

So without a complete With turnkey solutions available, companies must have a good understanding of how to make MLOps effective in implementing all aspects of AI. That means building expertise broadly, said Meagan Gentry, national practice manager for the AI ​​team at technology consultancy Insight.

MLOps covers the entire scope from data collection, verification and analysis, to managing machine resources and tracking model performance. There are many auxiliary tools that can be deployed locally, in the cloud or at the edge. Some of these tools are open source and some are proprietary. of.

But mastering technology is only one aspect, MLOps also draws on the agile methods and principles of iterative development from DevOps, Gentry said. Additionally, as with any agile-related field, communication is crucial.

"Communication in each role is very important, communication between data scientists and data engineers, communication with DevOps, and communication with the entire IT team."

For just For starting companies, MLOps can be confusing. There are many general principles, dozens of related vendors, and even a lot of open source tool sets.

"There are all kinds of pitfalls here," said Helen Ristov, senior manager of enterprise architecture at Capgemini Americas. "Many of them are still under development, and there is no formal set of guidelines yet. Like DevOps, this is still an emerging technology, and it will take some time for guidelines and related policies to be rolled out."

Ristov It is recommended that enterprises should start their MLOps journey with a data platform. "Maybe they have data sets, but they are in different places and there is not a unified environment."

Enterprises don't need to move all their data to one platform, but they do need one, she said. Methods introduce data from different data sources, and different applications have different situations. For example, data lakes are ideal for businesses that perform large amounts of analysis at high frequency and low-cost storage. MLOps platforms often have tools for building and managing data pipelines and tracking different versions of training data, but this is not a one-size-fits-all approach. Then there are other aspects like model creation, version control, logging, measuring feature sets, managing the model itself, and more.

"There's a lot of coding involved," Ristov said. Building an MLOps platform can take months, and platform vendors still have a lot of work to do when it comes to integrations.

"There is a lot of room for development in these different directions, many tools are still being developed, the ecosystem is very large, and people are just picking and choosing what they need. MLOps is still in its 'adolescence', and most enterprises Organizations are still looking for the most ideal configuration."

Market Landscape of MLOps

IDC's Subramanian said the MLOps market size is expected to grow from $185 million in 2020 to approximately $700 million in 2025 USD, but it’s also possible that this market is significantly undervalued because MLOps products are often bundled with larger platforms. He said the true size of the MLOps market could exceed $2 billion by 2025.

Subramanian said that MLOps vendors tend to fall into three broad categories. The first are large cloud providers, such as AWS, Azure and Google Cloud, which provide MLOps functionality as a service to customers.

The second category is machine learning platform manufacturers, such as DataRobot, Dataiku, Iguazio, etc.

"The third category is what we used to call data management vendors, such as Cloudera, SAS, DataBricks, etc. Their advantages lie in data management capabilities and data operations, and then extend to machine learning capabilities, and ultimately to MLOps Capabilities."

Subramanian said that these three areas have shown explosive growth, and what will make MLOps vendors stand out is whether they can support both local environments and cloud deployment models, and whether they can implement trusted, Responsible artificial intelligence, whether it is plug-and-play and whether it is easy to expand, this is the aspect that reflects the difference. ”

According to a recent IDC survey, the lack of various methods to implement responsible AI is one of the top three barriers to the spread of artificial intelligence and machine learning, tied for second place with the lack of MLOps. Caused by This situation is largely due to the fact that there is no other choice but to adopt MLOps, said Sumit Agarwal, research analyst for artificial intelligence and machine learning at Gartner.

"The other methods are manual, so, there is really no Other options were available. If you want to scale, you need automation. You need code, data, and model traceability. "

According to a recent survey by Gartner, the average time it takes for a model to go from proof of concept to production has shortened from 9 months to 7.3 months. "But 7.3 months is still a long time, and enterprises There are many opportunities for organizations to leverage MLOps. "

The corporate culture change brought about by MLOps

Amaresh Tripathy, global head of analytics at Genpact, said that implementing MLOps also requires a cultural change as an enterprise AI team.

“A data scientist often comes across as a mad scientist trying to find a needle in a haystack. But in reality data scientists are discoverers and explorers, not factories producing widgets. "Enterprises often underestimate the effort they need to make.

"People can better understand engineering and have such and such requirements for user experience, but for some reason, people have completely different requirements for deployment models. One would assume that all data scientists who are good at testing environments will naturally deploy these models, or can send a few IT staff to deploy them, which is wrong. People don’t understand what they need. ”

Many companies are not aware of the knock-on effects MLOps may have on other aspects of the company, which often leads to huge changes within the company.

“You can put MLOps in call centers and the average response time will actually increase because simple things are left to machines and AI to handle, and things that are left to humans actually take longer time, because these things tend to be more complex. So you need to rethink what these jobs are, what kind of people you need, what kind of skills these people should have."

Tripathy said that today, a company Less than 5% of decisions in organizations are driven by algorithms, but this is changing rapidly. "We predict that in the next five years, 20% to 25% of decisions will be driven by algorithms, and every statistic we see shows that we are at an inflection point in the rapid expansion of artificial intelligence."

He believes that MLOps is a critical part. Without MLOps, you can't use AI consistently. MLOps is the catalyst for scaling enterprise AI.

The above is the detailed content of To accelerate AI development, how can companies use MLOps to improve production efficiency?. For more information, please follow other related articles on the PHP Chinese website!