Those who have gained some experience with machine learning, trying to digest the huge volumes of data that make it meaningful, quickly get to grappling with algorithms needed to optimize how the data is processed. And no matter how brilliant is the AI worker or data scientist trying to tame the data, it’s usually a laborious, trial-and-error and often expensive process.
That’s what Scott Clark observed when he was a student at Cornell University, and later when he worked on the ad targeting team at Yelp. To help himself, he worked on some tools to put a degree of automation into the process of optimizing data used for machine learning. Now a PhD in applied mathematics, Scott is the CEO and co-founder of SigOpt, in the business of helping companies get to machine learning in a more practical way.
“We were working in bioinformatics problems,” Clark said of his Cornell days. “It was like trying to solve a huge jigsaw puzzle with a supercomputer, taking things apart and putting them back together. It’s a really difficult computational problem; a lot of expertise is required to solve it. The algorithms you need have many adjustments to be made. It was often left to the grad students how to figure out how to configure them.”
“It was a trial and error process, very painstaking, and very expensive because it required so much supercomputer time,” he said. He observed other grad students in other departments facing similar problems and taking similar approaches, so he started searching for a better way. “I came across a guy studying optimal learning, trying to tackle the problem head on, of how to onfigure a time-consuming and expensive system as quickly as possible. “
He wound up concentrating his PhD studies on this field of optimizing machine learning, then he spent two-and-a-half years at Yelp configuring the ad system. “They had similar problems. If you configured it just right, you made a lot of money. But it took time and was expensive.”
About three years ago, he open-sourced some of the code he had devised; it became popular, but it still required domain expertise support. So he decided to form SigOpt with co-founder Patrick Hayes. They have since received several rounds of financing, $8 million total, and have a team of 14 optimization and optimal learning experts in San Francisco. Most of the customers are in oil and gas, financial services and academia, but anyone trying complex machine learning is a target. “We are trying to unlock this black box of parameter tuning,” Clark said.
For a demonstration example, SipOpt compared their own method run on an NVIDIA GPU, with a random search and grid search approaches run on a standard infrastructure. The results showed the SipOpt approach to have a lower cost and greater accuracy. Using SigOpt and the GPUs, results were obtained with $11 of computing costs. Using random search on a standard infrastructure was over 400 times more expensive and took more time.
“Saving money on your AWS bill can add up quickly,” Clark said. “Our customers can spend more time building models than tuning them.”
Clark recently spoke at AWS AI Day, in a talk on setting up your first neural network. More than half the questions were on how to pick the parameters, and often the suggestions were to “use brute force” or “do it in your head,” which he called “unsatisfying,” noting, “No human is good at doing 15-dimension observations in their head.”
Writing in The Next Platform, analyst Nicole Hemsoth stated, “The grand challenge for SigOpt is to generalize parameter tuning across a wide range of machine learning applications.”
– By John P. Desmond, AI Trends Editor
Get more information at SigOpt.com.