Enterprises are leveraging machine learning operations to gain business benefits

​When companies first begin deploying AI and launching machine learning projects, the focus is often on a theoretical level. Is there a model that can provide the necessary results? How to build this model? How to train this model?

But the tools data scientists use to develop these proofs of concept often cannot Translates well to production systems. As a result, it takes more than nine months on average to deploy an AI or machine learning solution, according to IDC.

IDC analyst Sriram Subramanian said: "We call it 'model speed,' which is the time it takes for a model to get from start to finish. Time.”

This is where MLOps comes into play. MLOps (Machine Learning Operations) is a set of best practices, frameworks and tools that help enterprises manage data, models, deployment, monitoring and other aspects of taking a theoretical proof of concept and putting an AI system into use.

“MLOps can reduce model speed to weeks—sometimes days,” Subramanian said. "Just like using development operations (DevOps) reduces the average time to develop an application, so you need to use MLOps."

By using MLOps, enterprises can build more models, innovate faster, and address more use cases. "The value proposition is very clear," he said.

IDC predicts that by 2024, 60% of enterprises will use MLOps to implement their machine learning workflows. Subramanian said that when companies are surveyed on the challenges they face when using AI and machine learning technologies, a lack of MLOps has emerged as a major barrier to enterprise adoption of AI and machine learning technologies, second only to cost.

Here, we’ll look at what MLOPs are, how they’re evolving, and what organizations need to use and keep in mind to make the most of this emerging approach to implementing AI technologies.

The evolution of MLOps

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

Today, MLOps provides a fairly powerful framework for implementing AI technology, said Zuccarelli, now an innovation data scientist at CVS Health. As an example, Zuccarelli pointed to a previous project he worked on to develop an app that could predict adverse outcomes, such as hospital readmission or disease progression.

“We were looking at data sets and models and talking to doctors to figure out the characteristics of the best models,” he said. "But to make these models truly useful, we need to put these models in front of actual users."

That means developing a mobile app that is reliable, fast and stable, And on the back end there is a machine learning system connected via API. “Without MLOps, we wouldn’t be able to guarantee that,” he said.

His team created a health dashboard for the model using the H2O MLOps platform and other tools. "You don't want the model to change significantly," he said. “And you don’t want to introduce bias. This health dashboard allows us to understand if changes have occurred in the system.”

Using the MLOps platform also allows us to make updates to production systems. "It's very difficult to replace a file without stopping the application from running," Zuccarelli said. "Even if the system is in production, MLOps tools can replace the system with minimal interference to the system itself."

He said that as the MLOps platform matures , these platforms will speed up the entire model development process, because companies do not have to do some repetitive work in vain for each project. Data pipeline management capabilities are also critical to implementing AI.

“MLOps comes into play if we have multiple data sources that need to communicate with each other,” he said. “You want all the data that flows into your machine learning model to be consistent and of high quality. As they say, garbage in, garbage out. If the information the model is fed is of poor quality, then its predictions themselves will be Very poor.”

The Fundamentals of MLOps: Moving the Target

But don’t assume just because there are platforms and tools now available that you can Ignore the core principles of MLOps. Businesses just starting out in this space should remember that at its core, MLOps is about creating a strong connection between data science and data engineering.

“To ensure that MLOps projects are successful, you need to have both data engineers and data scientists on the same team,” Zuccarelli said.

In addition, equipping the platform with the necessary tools to prevent bias, ensure transparency, provide explainability and support ethics – tools that are still under development, he said. “It definitely still requires a lot of work because it’s such a new area.”

So if there isn’t a complete turnkey solution available, businesses have to be proficient All aspects make MLOps so efficient when implementing AI technology. That means learning expertise on the job, said Meagan Gentry, domestic practice manager for the AI ​​team at Tempe-based Insight technology consulting firm.

MLOps covers everything from data collection, validation and analysis to managing machine resources and tracking model performance. Some of the tools that help businesses can be deployed on-premises, in the cloud, or at the edge. These tools can be open source or proprietary.

But technical knowledge is only part of the solution. MLOps also draws on agile methods from development operations (DevOps) and iterative development principles, Gentry said. Additionally, as in areas related to agile development, communication is crucial.

“Communication between every character is critical,” she said. "Communication between data scientists and data engineers. Communication with DevOps, and communication with the larger IT team."

For companies just starting out , MLOps can be confusing to you. Some general principles can be seen, there are dozens of vendors, and there are even more open source toolboxes.

"There are some pitfalls," said Helen Ristov, senior manager of enterprise architecture at Capgemini Americas. "A lot of these pitfalls are in the development process. There's not a formal set of guidelines like you see with DevOps. It's an emerging technology and some of the guidelines and strategies will take some time to develop. ”

Ristov recommends that companies start their MLOps journey from their data platform. "Maybe they have multiple data sets, but they're in different places and not in a very cohesive environment," she said.

She said enterprises do not need to move all data to one platform, but they do need a way to bring in data from different data sources, which may vary depending on the application. difference. For example, a data lake is a good fit for companies that require low-cost storage and frequently perform large amounts of analysis.

She said that MLOps platforms usually provide tools to build and manage data pipelines while recording different versions of training data, but this is not a one-and-done solution.

Then also provides model creation, version management, logging, measurement feature sets, and other aspects of managing the model itself.

"It involves a lot of coding," Ristov said, adding that setting up an MLOps platform can take months, and when it comes to integration While working, platform providers still have a lot of work to do.

"There's a lot of growth going in different directions," she said. "There are a lot of tools being developed, the ecosystem is very large, and people are just picking and choosing what they need. MLOps is in its immature stage. Most organizations are still figuring out the best configuration."

Understanding the MLOps landscape

The MLOps market size is expected to grow from approximately $185 million in 2020 by 2025, according to IDC’s Subramanian to about $700 million. But he said that could be a serious underestimate because MLOps products are often bundled with larger platforms. The true size of the market could exceed $2 billion by 2025, he said.

Subramanian said MLOps vendors generally fall into three categories, starting with the large cloud providers, including Amazon Web Services (AWS), Azure Cloud and Google Cloud; These cloud platforms provide MLOps functionality as a service.

Then there are some machine learning platform vendors, such as DataRobot, Dataiku, Iguazio, etc.

"The third category is what they used to call data management vendors," he said. "Companies like Cloudera, SAS, and DataBricks. Their strength is in data management capabilities and data operations, and then they expand to have machine learning capabilities, and ultimately to MLOps capabilities."

Subramanian said all three areas are seeing explosive growth, adding that the key to making MLOps vendors stand out is whether they can support deploying models both on-premises and in the cloud, and whether they can implement trustworthy and responsible AI, whether they can provide plug-and-play solutions, and whether their solutions can be easily expanded. "That's where the differences between vendors come in," he said.

According to a recent survey by IDC, a lack of methods for implementing responsible AI is one of the top three barriers to using AI and machine learning technologies, tied for second place with the lack of MLOps itself. .

Sumit Agarwal, research analyst for AI and machine learning technologies at consulting firm Gartner, said this is largely because there is no alternative to using MLOps.

"Every other method is manual," he said. "So, there's really no other choice. If you want to scale, you need automation. You need traceability of your code, data and models."

According to consulting firm Gartner According to a recent survey, the average time it takes for a model to go from proof of concept to production has dropped from 9 months to 7.3 months. “But 7.3 months is still a long cycle,” Agarwal said. “There are many opportunities for organizations to leverage MLOps.”

Shifting Organizational Culture to MLOps

Genpact Global Analytics Business MLOps also requires a change in organizational culture on a company's AI team, said principal Amaresh Tripathy.

“The common image people have of a data scientist is that of a mad scientist trying to find a needle in a haystack,” he said. "A data scientist is a discoverer and explorer, not a factory floor churning out widgets. But that's what you need to do when you really want to scale."

He Says companies tend to underestimate the effort they need to expend.

"People have a better understanding of software engineering," he said. “There are a lot of rules about user experience and requirements. But somehow people don’t think they have to go through the same process when they deploy a model. There’s also a misconception that all data scientists who are good at working in a test environment People will naturally deploy and be able to deploy a certain model, or they can send a few IT colleagues to do it. There is a lack of understanding of what they need to do."

Businesses have yet to realize that MLOps can have a knock-on effect on other parts of the company, often resulting in dramatic changes.

“You can deploy MLOps in the customer service center, but the average response time will actually increase because some simple tasks are handled by machines and AI and handed over to humans. jobs actually take longer because they are more complex," he said. "So you need to rethink what the work is going to be, what kind of people you need, and what kind of skills they should have."

He said that today, in an organization, Less than 5% of decisions are driven by algorithms, but this is changing rapidly. "We predict that over the next five years, 20 to 25 percent of decisions will be driven by algorithms. Every statistic we examine shows that we are at an inflection point for the rapid expansion of AI."

MLOps is a key element, he said.

"One hundred percent," he said. “You can’t sustainably use AI without MLOps. MLOps are a catalyst for expanding the use of AI in the enterprise.”

The above is the detailed content of Enterprises are leveraging machine learning operations to gain business benefits. For more information, please follow other related articles on the PHP Chinese website!