After years-long collaboration between the University and Google, a new Google AI lab is set to open next week at 1 Palmer Square in the town of Princeton. The lab, headed by computer science professors Elad Hazan and Yoram Singer, will continue research on the optimization of machine learning techniques for speed and accuracy.
The professors’ theoretical work has diverse applications, from self-driving cars to facial recognition technology.
“It’s an exciting day and age that we are achieving learning on a large scale for some very important problems,” Hazan said.
The Google AI lab began with a startup co-founded by Hazan and Jacob Abernethy, a professor of computer science at Georgia Tech. Called In8 (pronounced “innate”), the startup was concerned with applications in robotics — how to control machinery like self-driving cars, drones, and robotic arms.
The group gathered significant interest from companies such as Amazon and Google. This past summer, In8 was acquired by Google, paving the way for the creation of the Google AI lab.
The lab will be a new hub for the small yet diverse group of graduate students, undergraduates, and full-time Google employees currently split between Hazan’s lab in Palmer Square and Google’s offices in New York.
This collaboration merges theoretical and applied research, where fundamental work in computer science and mathematics can translate to practical uses and technologies.
“Google provides an excellent opportunity to apply the theory, develop it further, and be exposed to real-world problems,” Hazan said.
Furthermore, the lab offers unique resources and computing power to run large-scale experiments.
“With Google’s resources, we have access to more interesting problems. We hear about what people need in deep learning, what problems they have, what kind of trade-offs they want. With these new problems in mind, we can come up with more impactful work,” Xinyi Chen ’17 said.
Chen, who was advised by Hazan while an undergraduate, is now a software engineer employed at Google.
“One big challenge in machine learning is defining a proper problem,” Chen said of her daily work in the lab. “It’s a lot of meetings and discussions. We explore different potential solutions.”
While some on the team, such as Chen, are employed full-time, others work part-time as research interns.
Karan Singh GS said he believes that the team will learn from working with those who put theoretical models into practice.
“Working in conjunction with people who actually practice machine learning brings out interesting theoretical challenges,” Singh said. “They make you question your models.”
Singh, along with Cyril Zhang GS and Brian Bullins GS, are part-time student researchers at Google while they complete their graduate work under Hazan.
While the focus of the Google AI lab will be the theoretical basis of reinforcement learning and large-scale optimization problems, the field is flexible and continues to grow, Zhang said.
The resources available at Google give the group access to more applied problems, such as natural language processing. By learning from streams of text, this technology can predict and suggest responses to emails, text messages, and more.
“The mission of machine learning is to design agents that are able to act intelligently in changing environments” in which there is noisy or incomplete information, Zhang said. “We’re trying to tackle the most fundamental mathematical abstractions of decision-making.”
For Bullins, the partnership with Google allows him to transfer theoretical models developed on a smaller-scale.
“It’s this blending of interesting theoretical foundations with real-world, incredibly large-scale applications,” accessible only at Google, Bullins said, that enriches his current work.
Still other members of the team are undergraduates already participating in cutting-edge research.
For instance, while most of the group focuses on the theoretical side of machine learning, Abby Van Soest ’19 works on experimentation, implementing efficient algorithms for exploration in different environments. This exploration has applications for robotics, such as how to get a robot to visit every area in a room if it is put in the center without any information. Van Soest, although not currently employed with Google, co-authored a paper with the team last year.
“I hoped that my skills as a programmer or a more applied researcher would be useful to them,” Van Soest said. She said she appreciated collaborating with theorists because “you learn a lot from them. It’s amazing what can be done.”
Although concerns have been raised about potential conflicts of interest, Hazan believes that partnership between universities and industries is essential for scientific progress.
“This is very positive, and encouraged” by parties in academia, industry, and government, Hazan said, emphasizing that the current initiative is neither unique nor new to Google AI.
As they transition to their new offices, the members of the Google AI team hope to continue their transformative work in machine learning optimization.
“Our goal is both scientific — to contribute to the understanding of some of the most intriguing questions ever asked about the nature of learning and intelligence — as well as applicative — to develop new methods that will positively impact society,” Hazan said. “It is a passion for us.”