Peter_Frazier_005_edit.jpg

Whether considering the cost of switching to something new or data mining other complex questions, machine learning crunches deep data for answers.
Dave Burbank
Dave Burbank

Peter_Frazier_077_edit.jpg

“This sort of question originally came up at Uber, but then I noticed it in other places too: academics choosing whether to start doing research in a hot new area, sellers on eBay choosing whether to start selling a new kind of fad product.”
Dave Burbank
Dave Burbank

PeterFrazierCollage_edit.jpg

Peter Frazier and his lab are continuously producing algorithms and mathematical modeling for machine learning and deep neural networks, applying them to answering tough questions and decision making.
Beatrice Jin; Dave Burbank
Beatrice Jin; Dave Burbank

Peter_Frazier_039_edit.jpg

“Deep neural networks are used for predicting all kinds of things…hedge funds for predicting movements of stocks…behavior of people on the web…autonomous vehicles for…figuring out whether something is a tree or a stop sign.”
Dave Burbank
Dave Burbank

Peter_Frazier_064_edit.jpg

Speaking about ORIE, “It’s an excuse to learn about whatever seems interesting at the moment, all different application domains. I can learn about chemistry and airplanes and autonomous vehicles…It’s pretty fun.”
Dave Burbank
Dave Burbank

How We Use Machine Learning

by Jackie Swift

If you’ve ever used Uber, then you probably know about surge pricing, times when the fare price suddenly increases for rides originating in a certain location. For riders, a surge can be an unanticipated expense that makes them look for other transportation. For drivers, it represents an opportunity: get to where the surge is happening and make more money.

“If you’re a driver and you’re far away from the area that is surging, you might consider going there if you can be sure the surge price will remain in effect,” says Peter I. Frazier, Operations Research and Information Engineering. “But you also know other drivers might get there before you and take all the rides, so you might decide not to go.”

To Switch or Not

Maybe just enough drivers arrive at the surging location to pick up all the passengers who need a ride. Maybe too few show up, and riders are left waiting in vain. Or perhaps so many drivers decide to follow the surge that some do not get a trip. What can be done to make sure things work well consistently? Frazier, who holds a half-time position at Cornell and also works for Uber as a staff data scientist, is currently researching this supply-and-demand conundrum as a Cornell scholar.

“This sort of question originally came up at Uber, but then I noticed it in other places too: academics choosing whether to start doing research in a hot new area, sellers on eBay choosing whether to start selling a new kind of fad product. Anything where there’s a cost of switching into a new area and an uncertain reward that depends on the number of other people that switch.” Joining with Cornell colleague Krishnamurthy Iyer and graduate student Pu Yang, PhD’19, Operations Research and Information Engineering, Frazier is analyzing data and building mathematical models to explore the phenomenon.

Along with the research he does for and about Uber, Frazier also works with many different collaborators, such as biochemists and material scientists, on a wide range of problems requiring machine learning, a type of artificial intelligence application that uses data and previous experience to adapt and learn. Frazier’s expertise helps narrow down decisions on what kind of data his colleagues should collect for a given purpose.

Searching for Peptides with a Special Enzymatic Activity

In one such project, Frazier joined with a number of collaborators at the University of California, San Diego (UCSD) and Northwestern University. The researchers wanted to find short peptides with a specific type of enzymatic activity that allows them to attach or detach molecules to proteins, making them prime candidates for applications such as drug delivery.

“When the people at UCSD approached me, they had examples of longer peptides that had this particular enzymatic activity, but they wanted shorter ones that had the same property,” Frazier says. “Initially, they had experimental data for 15 peptides. So we went through rounds where my algorithm would recommend roughly 500 peptides based on the earlier experimental data. We’d test all 500, and then using the data from those 500, we’d run through the algorithm again and get another 500 and test those, and so on.”

Using this optimal learning approach, the researchers ultimately developed a method to identify short peptides as selective substrates for enzymes. “What my algorithm does is something humans also can do pretty well for a few things at a time,” says Frazier. “A human could probably keep the data on three peptides in their head and judge each peptide on the relevance to the task of being biologically active but also diverse. But to do that simultaneously for 500 peptides would be really tough.”

Neural Networks for Predicting the Behavior of Stocks, Web Browsing, Autonomous Vehicles, Much More

Frazier also works on projects dealing with deep neural networks, systems of software patterned after neurons in the brain, with many layers. Neural networks use sophisticated mathematical modeling to process data in complex ways. “Deep neural networks are used for predicting all kinds of things,” he says. “They’re used a lot in hedge funds for predicting movements of stocks. They’re used to predict the behavior of people on the web, and they’re used in autonomous vehicles for things like figuring out whether something is a tree or a stop sign. But they require a lot of fine tuning to get them to work well.”

“Then you’d push a button and wait six hours while the program processes the pictures. If you set your parameters well…the program will be able to tell you with 95 percent accuracy whether a picture contains a stop sign or not.”

To train a deep neural network to accurately recognize a stop sign, for instance, you might load a computer program with 100 thousand images; 50 thousand would contain a stop sign and 50 thousand would not contain one. “Then you’d push a button and wait six hours while the program processes the pictures,” Frazier says. “If you set your parameters well, at the end of that time, the program will be able to tell you with 95 percent accuracy whether a picture contains a stop sign or not.”

Figuring out how to set the parameters, however, isn’t all that easy. “You may run your program with 100 thousand pictures for six hours on your supercomputer and realize when it’s done, it doesn’t actually work very well at predicting whether something is a stop sign or not,” Frazier says. “So you try new parameters and run it again. You could spend two weeks doing that. To reduce the time and effort, I create algorithms that can choose the parameters automatically faster than a human would be able to do.” Frazier has applied some of his deep neural network algorithms to applications as varied as finding the best website design to increase user traffic and finding the best shape of an airplane wing to optimize a plane’s fuel efficiency.

The Operations Research and Information Engineering Advantage

Frazier decided to pursue operations research and information engineering because he wanted to work on something applicable to everyday questions that would help people and also satisfy his curiosity about the world around him.

“It’s an excuse to learn about whatever seems interesting at the moment, all different application domains,” he says. “I can learn about chemistry and airplanes and autonomous vehicles, and then after a couple of years, I can go do something else. It’s pretty fun.”