The Machine Learning Building Blocks Developers Require to do ‘MLOps’

It is becoming increasingly apparent that enterprises must, in the near future, integrate machine learning (ML) in their software development pipelines in order to remain competitive in some way. But what exactly does AI and ML integration mean for any organization, whether a small-to-medium-sized company or a large multinational or a government agency with small offices?

In the recently recorded “Advanced Developer Workloads with Built-In AI Acceleration” livestream broadcast, Intel’s AI expert Jordan Plawner, director of products and business strategy, artificial intelligence, described the building blocks DevOps teams require to harness the power of AI and ML for application development and deployments. Hosted by Alex Williams, founder and publisher of The New Stack, Plawner was able to draw upon his experience to explain how data scientists and DevOps teams can take advantage of ML. Since his main role is to ensure Intels’ Xeon server processors and software are built to spec for DevOps teams’ workloads and library needs, Plawner is able to draw on his deep understanding of MLOps hardware and supporting software requirements.

“We are making it easy for developers and DevOps people to stand up and run AI-based workloads on our general-purpose processor Xeon,” Plawner said. “So, I’m responsible for the roadmap, and all the kind of software capabilities that we deploy for developers.”

The AI/ML data challenge: @intel‘s #JordanPlawner: Data science workstations are a huge answer to manage massive amounts of data..Data clusters that bring the compute and the data together are part of the solution. #sponsored #ai @alexwilliams https://t.co/SBJ4AdQ0TF pic.twitter.com/CMxl7pv6de

— The New Stack (@thenewstack) April 21, 2021

At the end of the day, machine learning (MLOps) processes are somewhat comparable to those of DevOps processes, while also diverging in significant ways. ML is, to begin with, non-binary, as systems are taught to draw inferences and perform analyses in ways that can often mimic that of the human brain. ML, in this way, is significantly more complex than simple “if-then” computing, even when compared to software used to automate processes that might also fall under the AI computing category.

As an example, a continuous integration/continuous delivery (CI/CD) ML pipeline with TensorFlow — an open source platform that Google uses for its ML application-development needs — must first be “taught” how to process often enormously complex calculations. Once taught, the application can then be applied to run ML applications. Google also relies on the Intel oneAPI framework for its use of TensorFlow.

MLOps “is really taking the idea of DevOps and applying it to the deep learning space in the AI space to help automate a lot of these things that developers shouldn’t be spending their time doing,” Plawner said. “Data scientists should be experimenting and the DevOps people need good MLOps tools so that they can help orchestrate and manage that environment.”

@intel “takes an expansive view of what is AI because customers have all kinds of different datasets..and need insights…while there are different “points of view” about “what is AI and what is not AI,” says #JordanPlawner. @thenewstack #Sponsored https://t.co/i4lINA78Fn

— BC Gain (@bcamerongain) April 21, 2021

The ML Mantra

Reflecting Intel’s philosophy, Plawner said “we take a very broad view of artificial intelligence.”

“MLOps is a set of very extensive techniques to allow you to gain insight from different kinds of data historically,” he said.

DevOps teams have become “used to seeing very structured data in rows and columns and databases but we now know that we also live in a sort of analog-to-digital world where we have lots of video, images and speech” that ML can translate and make recommendations, Plawner said.

ML process are often required to draw inferences and make actionable recommendations for data that historically “has been very, very hard to gain insights from what we call ‘non-structured data.’” These datasets might include data images that might not have a regular pattern or structure from which a non-ML-assisted application would be able to draw inferences.

“What I think about AI is the machine-learning and deep-learning techniques, and all the end-to-end pipeline data management needed in order to gain insights from all the data that is collected,” Plawner said.

Software remains key and Python rocks. @intel‘s #JordanPlawner: There are many “correlating improvements in software development, especially with Python.. the most commonly used language…it’s what developers go to. #sponsored #ai @alexwilliams https://t.co/SBJ4AdQ0TF pic.twitter.com/whLphyTCUg

— The New Stack (@thenewstack) April 21, 2021

Virtualization and containers environments with libraries — supported mainly with Python — should increasingly be used to create the ML pipeline so that data scientists can worry less about operations and concentrate on their development work for ML applications.

IT mst @intel‘s #JordanPlawner: “One thing that kills projects in a deep learning space, and most deep learning projects still do not go into production..they get stuck in the whole data lake flood of how to manage the data.” #sponsored #ai @alexwilliams https://t.co/SBJ4AdQ0TF pic.twitter.com/xVgJ7l9365

— The New Stack (@thenewstack) April 21, 2021

“I’ll tell the developer, ‘okay if you want to read the white papers about all these innovations, that’s great. But if you just want the latest innovations, point and click on this container and download it…and just run it on the latest Intel processor,’” Plawner said. “So, I think really that’s what we’re talking about: that abstraction, as we just want to pile all this into a container and make it much easier to run.”

View the whole discussion here.