The future of Machine Learning (ML) is more: more projects, more stakeholders, and more migrations of high-performance computing (HPC) for Machine Learning workloads to the Hybrid Cloud.
A recent survey conducted by Dimensional Research looked at the market momentum of Machine Learning, why companies are moving in that direction, and the hurdles they must overcome to take advantage of Hybrid Clouds for their projects.
Among its findings are that 69 percent of 344 participants in the survey from across the world have three or more teams requesting Machine Learning projects, and 36 percent already have 10 or more researchers within ML teams. Ninety-six percent of them stated that their ML projects will grow over the next two years and 88 percent plan to use the Hybrid Cloud for their work.
Diversity is the name of the game for ML projects. The five top Machine Learning projects the respondents said they deployed were: IT optimization, security, fraud detection, natural language, and customer learning.
The main industries expected to take the lead in adopting Machine Learning were technology, financial services, healthcare, retail, telecommunications, and manufacturing. IT optimization will likely lead in deployments because the technology industry naturally is cognizant of that requirement, while security and financial services make for a strong pairing. HPC lets companies undertake these and other use cases — product design, analysis, testing — and do them faster.
“The fact that ML is really popular and ready to grow is not a surprise,” said Gary Tyreman, president and CEO of Univa, which sponsored the survey and offers solutions for workload orchestration for high-scale and high-performance environments.
Nor was it surprising to see 82 percent of respondents looking to leverage Hybrid Cloud infrastructure, given that a large majority (71 percent) already or expect to use graphic processer units (GPUs) for high-performance Machine Learning. “GPUs are very expensive, and they iterate quickly,” Tyreman explained. The more GPU iterations a company does for in-house HPC requirements, the more money it winds up spending on infrastructure rather than Data Science.
“What was surprising,” Tyreman said, “is that the percentage of projects in production now was lower than expected” — just 22 percent. Half of respondents are just at the stage of researching potential projects.
It’s not because companies don’t want to have more ML projects in production, said Rob Lalonde, Univa Vice President and General Manager of Navops, the company’s product suite for migrating HPC workloads to the Vloud.
Contributing to the low deployments, he said, is the length of time it takes to master data movement and loads, not to mention finding skilled Data Scientists and personnel who can create training models and understand complicated data flows.
The survey shows that 63 percent of respondents pointed to lack of experience as a main issue for ML projects, with 37 percent noting that complicated data flows were part of the problem in moving ML projects into production. Having to migrate large databases also ranked high among the challenges.
HPC and the Hybrid Cloud
The technology is there, though, from Amazon, Google, Microsoft, and others to get a head start with Machine Learning. Amazon Machine Learning, Google TensorFlow, and Microsoft Cognitive Toolkit (CNTK) are leaders today, but results show that in two years the use of Amazon’s and Microsoft’s technology will significantly accelerate.
The Hybrid Cloud is certainly part of the technology picture. Breaking down the 82 percent of respondents looking to leverage Hybrid Cloud infrastructure, 44 percent of respondents are using the Hybrid Cloud today for Machine Learning projects, and 38 percent plan to use it in the future. “The Hybrid Cloud means you have additional capacity but also access to resources you might need that you didn’t have on premise,” Lalonde said. “It’s a big driver to go from CAPEX to OPEX, to the Cloud and access to the ML software stack.”
Migrating on-premise high-performance computing platforms to the Cloud facilitates ML operations, expanding capabilities while cutting costs. In fact, 88 percent of the respondents have HPC responsibilities, indicating a direct correlation between HPC and Machine Learning.
But the driver for Machine Learning in the first place is what the business wants to do. “Fraud detection is not driven by HPC staff but by someone in the business who wants to improve trading algorithms,” Lalonde said. “Automated driving is not driven by HPC but by the people in auto companies’ design operations and the forward-looking strategic people who are looking at next generation cars.”
Sixty-four percent of companies expect to use two or more Machine Learning tools going forward. “It comes down to there being a lot of new tools out there, so companies are trying different things,” Lalonde said. “I also think the tools out there today are designed to solve different problems and address different parts of the workflow.”
Univa Looks Forward
Univa’s strategy in all this revolves around helping companies undertake the expected and significant HPC Cloud migrations. From an infrastructure standpoint, it’s investing heavily in methodologies and processes, new products, and connectivity requirements.
“The reality to us is whether you are running an app that does computational fluid dynamics or regression analysis, the workload is the workload, and we sit at the operating system level,” Lalonde said.
Lalonde added that there’s no reason to force customers to digest only one way of getting to or even taking advantage of the Hybrid Cloud for Machine Learning.
“Often people say anything that should go to the Cloud, send it there. But we look at workloads, objects, and attributes and set up rules or routines. We can determine this job should go or can’t go. And if it should, under what constructs — for example, what will be the cost of moving data.”
According to Tyreman, its technology can help businesses increase utilization of their HPC computing platforms. “In the past, when they hit scale and also have peaks and valleys, they would likely over-provision,” he said. But extruding on-premise clusters to the Hybrid Cloud can increase HPC utilization to 80 percent to 90 percent, so companies get more value.
Image used under license from Shutterstock.com